[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nemo-mbridge-recipe-recommender":3,"mdc-9yc4mp-key":31,"related-repo-nvidia-nemo-mbridge-recipe-recommender":5454,"related-org-nvidia-nemo-mbridge-recipe-recommender":5560},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":20,"repoUrl":21,"updatedAt":22,"license":23,"forks":24,"topics":25,"repo":26,"sourceUrl":29,"mdContent":30},"nemo-mbridge-recipe-recommender","recommend Megatron Bridge training recipes","Recommend and customize Megatron Bridge recipes for a user's model, GPU count, and training goal. Indexes library recipes (pretrain\u002FSFT\u002FPEFT) and performance recipes.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19],{"name":13,"slug":14,"type":15},"Deep Learning","deep-learning","tag",{"name":17,"slug":18,"type":15},"Machine Learning","machine-learning",{"name":9,"slug":8,"type":15},2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:26:30.145791","Apache-2.0",281,[],{"repoUrl":21,"stars":20,"forks":24,"topics":27,"description":28},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fnemo-mbridge-recipe-recommender","---\nname: nemo-mbridge-recipe-recommender\nlicense: Apache-2.0\ndescription: Recommend and customize Megatron Bridge recipes for a user's model, GPU count, and training goal. Indexes library recipes (pretrain\u002FSFT\u002FPEFT) and performance recipes.\nwhen_to_use: User wants a starting recipe or training config; 'which recipe', 'recommend recipe', 'how to train Llama', 'starting config for X GPUs', 'what recipe for SFT'.\n---\n\n# Auto Recipe — Recipe Index & Recommendation\n\nThis skill indexes every shipped recipe and helps users pick the right starting\nconfig, adjust parallelism, and avoid common pitfalls.\n\n## How to Use This Skill\n\n1. Ask the user for: **model name\u002Fsize**, **GPU count & type**, **training goal**\n   (pretrain \u002F SFT \u002F PEFT), and **sequence length** (if non-default).\n2. Look up the best-match recipe in the index below.\n3. Recommend the recipe function name + entry-point command.\n4. Provide adjustment advice (parallelism resizing, batch tuning, pitfalls).\n\n## First Answer Checklist\n\nWhen recommending recipes, always include these distinctions before the long\nindex details:\n\n1. **Library recipes** under `src\u002Fmegatron\u002Fbridge\u002Frecipes\u002F` are for functional\n   training and use `scripts\u002Ftraining\u002Frun_recipe.py`.\n2. **Performance recipes** under `scripts\u002Fperformance\u002F` are for upper-bound\n   throughput benchmarks. They use mock data and should not be presented as\n   production training recipes.\n3. For a first-time Bridge smoke test, recommend `llama3_8b_sft_config` with\n   mock data via `--dataset llm-pretrain-mock`. Do not use `llm-finetune` for\n   the setup-only tryout unless the user specifically asks for an SFT data path.\n4. For normal SFT recommendations, use `--dataset llm-finetune`; for pretrain\n   and mock validation recommendations, use `--dataset llm-pretrain-mock`.\n5. After the recipe and dataset, give the required resizing rules: TP must\n   divide `num_key_value_heads`, keep TP within one node unless using\n   NVL72-class interconnect, enable SP when TP > 1, configure CP for long\n   context, DP is implicit, and reduce `micro_batch_size` first on OOM.\n\n---\n\n## Entry Points\n\n### Library recipes (functional training)\n\n```bash\n# Pretrain with mock data\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe \u003Crecipe_function_name> \\\n    --dataset llm-pretrain-mock\n\n# SFT with SQuAD\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe \u003Crecipe_function_name> \\\n    --dataset llm-finetune\n\n# Override any field via CLI\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe llama3_8b_pretrain_config \\\n    --dataset llm-pretrain-mock \\\n    'model.tensor_model_parallel_size=2' \\\n    'training.global_batch_size=64'\n```\n\n### Performance recipes (throughput benchmarks)\n\n```bash\npython scripts\u002Fperformance\u002Frun_script.py \\\n    --recipe \u003Cmodel_family> \\\n    --gpu_type h100 \\\n    --num_gpus 64 \\\n    --data mock\n```\n\nSee the Performance Recipe Index for important caveats before using these for anything beyond throughput benchmarking.\n\n---\n\n## Perf Recipe Layout\n\nPerformance recipes use the same **Python function** format as library recipes,\nbut live in a dedicated namespace for throughput benchmarking:\n\n- Perf recipes live in `src\u002Fmegatron\u002Fbridge\u002Fperf_recipes\u002F\u003Cfamily>\u002F\u003Chardware>\u002F\u003Cmodel>.py`\n- Each perf recipe is a **self-contained Python function** (e.g. `llama3_8b_pretrain_8gpu_h100_bf16_config()`)\n- Recipe names encode model, task, GPU count, hardware, precision, and optional variant\n- `scripts\u002Fperformance\u002Futils\u002Futils.py` derives compatibility `WorkloadBaseConfig` views from the flat recipe itself\n- Shared helpers: `_benchmark_common()` (50 iters, timing, TE RNG), `_perf_precision()` (bf16 \u002F fp8_cs \u002F fp8_mx \u002F nvfp4)\n\n**Why Python, not YAML?** Previous YAML-based approaches had problems:\nrecipe logic was split across multiple indirection layers, configs were not\nself-contained, and the two-level pipeline made maintenance and debugging\ndifficult. Python functions are explicit, greppable, and composable.\n\nThe training launcher can invoke both library recipes and perf recipes without\nthe removed legacy config package.\n\n---\n\n## Library Recipe Index\n\nAll recipes live under `src\u002Fmegatron\u002Fbridge\u002Frecipes\u002F`. Each function returns a\n`ConfigContainer` with model, training, optimizer, and data settings.\n\n### Llama\n\n| Recipe | Mode | TP | PP | CP | SP | GPUs (min) | Seq Len |\n|--------|------|----|----|----|----|------------|---------|\n| `llama2_7b_pretrain_config` | Pretrain | 2 | 1 | — | — | 2 | 4K |\n| `llama3_8b_pretrain_config` | Pretrain | 2 | 1 | — | ✓ | 2 | 8K |\n| `llama3_8b_16k_pretrain_config` | Pretrain | 2 | 1 | 2 | ✓ | 4 | 16K |\n| `llama3_8b_64k_pretrain_config` | Pretrain | 2 | 1 | 4 | ✓ | 8 | 64K |\n| `llama3_8b_128k_pretrain_config` | Pretrain | 2 | 1 | 8 | ✓ | 16 | 128K |\n| `llama3_70b_pretrain_config` | Pretrain | 8 | 4 | — | ✓ | 32 | 8K |\n| `llama3_70b_16k_pretrain_config` | Pretrain | 8 | 4 | 2 | ✓ | 64 | 16K |\n| `llama3_70b_64k_pretrain_config` | Pretrain | 8 | 4 | 4 | ✓ | 128 | 64K |\n| `llama31_405b_pretrain_config` | Pretrain | 8 | 16 | — | ✓ | 128 | 8K |\n| `llama3_8b_sft_config` | SFT | 2 | 1 | — | ✓ | 2 | 8K |\n| `llama3_70b_sft_config` | SFT | 4 | 4 | — | ✓ | 16 | 8K |\n| `llama31_405b_sft_config` | SFT | 8 | 8 | — | ✓ | 64 | 8K |\n| `llama3_8b_peft_config` | PEFT | 1 | 1 | — | — | 1 | 8K |\n| `llama3_70b_peft_config` | PEFT | 2 | 4 | — | ✓ | 8 | 8K |\n| `llama31_405b_peft_config` | PEFT | 4 | 8 | — | ✓ | 32 | 8K |\n\n### Qwen2 \u002F Qwen2.5\n\n| Recipe | Mode | TP | PP | Sizes |\n|--------|------|----|----|-------|\n| `qwen2_*_{pretrain,sft,peft}_config` | All | 1–8 | 1–4 | 500M, 1.5B, 7B, 14B, 32B, 72B |\n| `qwen25_*_{pretrain,sft,peft}_config` | All | 1–8 | 1–4 | 500M, 1.5B, 3B, 7B, 14B, 32B, 72B |\n\n### Qwen3 (Dense)\n\n| Recipe | Mode | TP | PP | CP | Sizes |\n|--------|------|----|----|-----|-------|\n| `qwen3_*_pretrain_config` | Pretrain | 1–8 | 1–2 | — | 600M–32B |\n| `qwen3_*_sft_config` | SFT | 1–8 | 1–2 | — | 600M–32B |\n| `qwen3_600m_sft_128k_config` | SFT | 1 | 1 | 8 | 600M (128K seq) |\n| `qwen3_*_peft_config` | PEFT | 1 | 1 | — | 600M–32B |\n\n### Qwen3 MoE\n\n| Recipe | Mode | TP | PP | EP | CP | GPUs |\n|--------|------|----|----|----|----|------|\n| `qwen3_30b_a3b_pretrain_config` | Pretrain | 1 | 1 | 8 | — | 8 |\n| `qwen3_30b_a3b_sft_config` | SFT | 1 | 1 | 8 | — | 8 |\n| `qwen3_30b_a3b_peft_config` | PEFT | 1 | 1 | 1 | — | 1 |\n| `qwen3_235b_a22b_pretrain_config` | Pretrain | 4 | 16 | 8 | 2 | 512+ |\n| `qwen3_235b_a22b_sft_config` | SFT | 4 | 8 | 8 | — | 256 |\n| `qwen3_235b_a22b_peft_config` | PEFT | 1 | 4 | 4 | — | 16 |\n\n### Qwen3-Next\n\n| Recipe | Mode | TP | PP | EP |\n|--------|------|----|----|-----|\n| `qwen3_next_80b_a3b_pretrain_config` | Pretrain | 1 | 4 | 8 |\n| `qwen3_next_80b_a3b_sft_config` | SFT | 1 | 2 | 8 |\n| `qwen3_next_80b_a3b_peft_config` | PEFT | 1 | 1 | 4 |\n\n### DeepSeek\n\n| Recipe | Mode | TP | PP | EP | GPUs |\n|--------|------|----|----|-----|------|\n| `deepseek_v2_lite_pretrain_config` | Pretrain | 1 | 1 | 8 | 8 |\n| `deepseek_v2_pretrain_config` | Pretrain | 1 | 4 | 32 | 128 |\n| `deepseek_v3_pretrain_config` | Pretrain | 2 | 16 | 64 | 2048 |\n| `deepseek_v3_pretrain_config_32nodes` | Pretrain | 2 | 8 | 32 | 256 |\n\n### GLM-4.5\n\n| Recipe | Mode | TP | PP | EP | GPUs |\n|--------|------|----|----|-----|------|\n| `glm45_355b_pretrain_config` | Pretrain | 2 | 8 | 16 | 256 |\n| `glm45_air_106b_pretrain_config` | Pretrain | 1 | 4 | 8 | 32 |\n| `glm45_355b_sft_config` | SFT | 2 | 8 | 16 | 256 |\n| `glm45_air_106b_sft_config` | SFT | 1 | 4 | 8 | 32 |\n| `glm45_355b_peft_config` | PEFT | 2 | 4 | 4 | 32 |\n| `glm45_air_106b_peft_config` | PEFT | 1 | 2 | 4 | 8 |\n\n### Gemma\n\n| Recipe | Mode | TP | PP | Sizes |\n|--------|------|----|----|-------|\n| `gemma2_*_{pretrain,sft,peft}_config` | All | 2–8 | 1–2 | 2B, 9B, 27B |\n| `gemma3_1b_{pretrain,sft,peft}_config` | All | 1 | 1 | 1B (32K seq) |\n\n### NemotronH \u002F Nemotron\n\n| Recipe | Mode | TP | PP | EP | Notes |\n|--------|------|----|----|-----|-------|\n| `nemotronh_{4b,8b,47b,56b}_*_config` | P\u002FS\u002FPEFT | 1–8 | 1–4 | — | Dense SSM-hybrid |\n| `nemotron_3_nano_*_config` | P\u002FS\u002FPEFT | varies | 1 | 8 | MoE + Mamba |\n| `nemotron_3_super_*_config` | P\u002FS\u002FPEFT | 4 | 1 | 8 | MoE + Mamba, ~40% CUDA graph gain |\n| `nemotron_nano_{9b,12b}_v2_*_config` | P\u002FS\u002FPEFT | varies | 1 | — | Dense |\n\n### Other Models\n\n| Recipe | Mode | Notes |\n|--------|------|-------|\n| `moonlight_16b_{pretrain,sft,peft}_config` | All | MoE EP=8 |\n| `olmoe_7b_{pretrain,sft,peft}_config` | All | MoE EP=8 |\n| `ministral3_{3b,8b,14b}_{sft,peft}_config` | SFT\u002FPEFT | Dense |\n| `gpt_oss_20b_*_config` | All | MoE + FP8\u002FMXFP8 variants |\n| `gpt_oss_120b_*_config` | All | MoE |\n| `vanilla_gpt_pretrain_config` | Pretrain | MLM\u002FBridge parity baseline |\n| `gpt3_175b_pretrain_config` | Pretrain | TP=4, PP=8, VP=6 |\n| `kimi_k2_pretrain_config` | Pretrain | 1T MoE, TP=2 PP=16 EP=32 |\n\n### VLM Recipes\n\n| Recipe | Mode | TP | PP | EP | GPUs |\n|--------|------|----|----|-----|------|\n| `gemma3_vl_{4b,12b,27b}_{sft,peft}_config` | SFT\u002FPEFT | 1–8 | 1–2 | — | 1–16 |\n| `qwen25_vl_{3b,7b,32b,72b}_{sft,peft}_config` | SFT\u002FPEFT | 1–8 | 1–4 | — | 1–32 |\n| `qwen3_vl_{8b,30b_a3b,235b_a22b}_{sft,peft}_config` | SFT\u002FPEFT | 1–4 | 1–8 | 1–32 | 1–512 |\n| `qwen35_vl_*_{sft,peft}_config` | SFT\u002FPEFT | varies | varies | varies | varies |\n| `glm_45v_{sft,peft}_config` | SFT\u002FPEFT | 1 | 8 | 4–16 | 64–512 |\n| `nemotron_nano_v2_vl_12b_{sft,peft}_config` | SFT\u002FPEFT | 2–4 | 1 | — | 8 |\n\n### Diffusion Recipes\n\n| Recipe | Mode | TP | CP |\n|--------|------|----|----|\n| `wan_1_3B_{pretrain,sft}_config` | P\u002FSFT | 1 | 8 |\n| `wan_14B_{pretrain,sft}_config` | P\u002FSFT | 2 | 4 |\n| `flux_12b_{pretrain,sft}_config` | P\u002FSFT | 2 | 1 |\n\n---\n\n## Performance Recipe Index\n\nPerf recipe source lives under `src\u002Fmegatron\u002Fbridge\u002Fperf_recipes\u002F`. The\nperformance launcher in `scripts\u002Fperformance\u002F` resolves those flat recipe names\nand derives compatibility workload views from the selected flat recipe when\nlegacy helper paths still need them.\n\n> **Important:** Perf recipes are designed for **upper-bound throughput\n> benchmarks**, not production training. They run **50 iterations** on **mock\n> data** by default. Throughput numbers are aspirational targets, not validated\n> convergence configs.\n\n### Llama 3 \u002F 3.1\n\n| Model | GPUs | GPU Types | Key Features |\n|-------|------|-----------|--------------|\n| Llama 3 8B | 8 | H100, B200, B300, GB200, GB300, R100 | CUDA graphs (local), FSDP on GB variants |\n| Llama 3 70B | 64 | H100, B200, B300, GB200, GB300 | TP comm overlap (userbuffers), FSDP, CUDA graphs |\n| Llama 3.1 405B | 128–1024 | H100, B200, B300, GB200, GB300 | TP+CP comm overlap (userbuffers), FSDP, heavy PP\u002FVP |\n\nSFT\u002FLoRA variants also exist (e.g. 8B SFT with packed sequences, 70B SFT on 32 GPUs).\n\n### DeepSeek V3\n\n| Model | GPUs | GPU Types | Key Features |\n|-------|------|-----------|--------------|\n| DeepSeek V3 (671B MoE) | 256–1024 | H100, B200, B300, GB200, GB300 | HybridEP dispatcher, MLA recompute, CUDA graphs (TE scoped) |\n\n### Qwen3 MoE\n\n| Model | GPUs | GPU Types | Key Features |\n|-------|------|-----------|--------------|\n| Qwen3 30B-A3B | 8–16 | H100, B200, B300, GB200, GB300 | MoE alltoall\u002Fflex dispatcher |\n| Qwen3 235B-A22B | 64–256 | H100, B200, B300, GB200, GB300 | TP comm overlap, CUDA graphs, MoE a2a overlap |\n| Qwen3-Next 80B-A3B | 64–128 | H100, B200, B300, GB200, GB300 | EP 64–128 |\n\n### Qwen3-VL\n\n| Model | GPUs | GPU Types | Key Features |\n|-------|------|-----------|--------------|\n| Qwen3-VL 30B-A3B | 8–16 | H100, B200, B300, GB200, GB300 | VLM + MoE |\n| Qwen3-VL 235B-A22B | 64–256 | H100, B200, B300, GB200, GB300 | VLM + MoE, TP comm overlap |\n\n### Kimi K2\n\n| Model | GPUs | GPU Types | Key Features |\n|-------|------|-----------|--------------|\n| Kimi K2 (1T MoE) | 256–1024 | H100, B200, B300, GB200, GB300 | Muon\u002FAdam optimizer, HybridEP, pipeline layout helpers |\n\n### NemotronH\n\n| Model | GPUs | GPU Types | Key Features |\n|-------|------|-----------|--------------|\n| Nemotron 3 Nano (30B MoE+Mamba) | 8–16 | H100, B200, B300, GB200, GB300 | TE CUDA graphs (attn+mamba+moe), HybridEP |\n| Nemotron 3 Super | 64 | H100, B200, B300, GB200, GB300 | TE CUDA graphs, EP=64 |\n| NemotronH 56B | 64 | H100, B200, B300 | TP=2–8, TE graphs (mamba+attn) |\n\n### GPT-OSS\n\n| Model | GPUs | GPU Types | Key Features |\n|-------|------|-----------|--------------|\n| GPT-OSS 120B | 64 | H100, B200, GB200 | EP=64, HybridEP on GB200 |\n\n---\n\n## Recommendation Decision Tree\n\n```text\nUser wants to train a model\n│\n├─ Know the model name?\n│   ├─ Yes → Look up in Library Recipe Index above\n│   │   ├─ Has a recipe for their size + mode? → Use it directly\n│   │   └─ No exact match? → Use closest size, adjust parallelism\n│   └─ No → Ask for model name, size, and HF model ID\n│\n├─ What's the training goal?\n│   ├─ Pretrain → Use *_pretrain_config\n│   ├─ SFT (full fine-tune) → Use *_sft_config\n│   └─ PEFT (LoRA\u002FDoRA) → Use *_peft_config (lowest GPU requirement)\n│\n├─ How many GPUs?\n│   ├─ 1 GPU → Only PEFT recipes work (TP=1, PP=1)\n│   ├─ 8 GPUs (1 node) → Most 8B–16B models, small MoE (EP=8)\n│   ├─ 16–64 GPUs → 70B dense, medium MoE\n│   └─ 128+ GPUs → 405B+, large MoE (DeepSeek V3, Kimi K2)\n│\n├─ Want throughput benchmarks?\n│   ├─ Yes → Use perf recipes (scripts\u002Fperformance\u002F)\n│   │   └─ ⚠️ These run on mock data for upper-bound perf only\n│   └─ No → Use library recipes (scripts\u002Ftraining\u002Frun_recipe.py)\n│\n└─ Long context?\n    ├─ > 8K → Need CP (context parallelism), check *_16k \u002F *_64k \u002F *_128k variants\n    └─ ≤ 8K → Default recipes work\n```\n\n---\n\n## Adjustment Advice (When Recommending)\n\n### Parallelism Resizing Rules\n\nWhen the user's GPU count differs from the recipe default:\n\n1. **TP must divide `num_key_value_heads`** (GQA constraint). E.g. if\n   `num_key_value_heads=8`, valid TP = {1, 2, 4, 8}.\n2. **TP should stay within a single node** (NVLink). TP > 8 requires\n   inter-node NVLink (e.g., GB200 NVL72).\n3. **PP adds pipeline bubbles.** Minimize PP; only increase when TP alone can't\n   fit the model. Use VP (virtual pipeline) to mitigate bubble overhead.\n4. **EP doesn't reduce dense-layer memory.** Only expert parameters shard with\n   EP. Shared attention\u002Fembeddings are replicated. For \"OOM with MoE\", increase\n   EP first, not TP.\n5. **SP should be True whenever TP > 1.** It eliminates redundant activation\n   copies and is essentially free.\n6. **CP requires all-to-all or ring attention.** Check `cp_comm_type`. For\n   GQA models, `a2a+p2p` hierarchical CP allows CP > num_kv_heads.\n7. **world_size = DP × TP × PP × CP × EP.** DP is implicit. Make sure the\n   product of explicit parallelisms divides your total GPU count.\n\n### Batch Size Tuning\n\n- Start with the recipe's `micro_batch_size`. If OOM, reduce to 1.\n- `global_batch_size` determines learning dynamics. Scale with DP:\n  `GBS = micro_batch_size × DP × gradient_accumulation_steps`.\n- For MoE, `micro_batch_size=1` is typical at scale.\n\n### Common Pitfalls to Warn About\n\n| Pitfall | Symptom | Fix |\n|---------|---------|-----|\n| TP > num_kv_heads | Crash: \"TP must divide num_query_groups\" | Reduce TP to a divisor of num_kv_heads |\n| PP without VP | Poor throughput (large bubble) | Set `virtual_pipeline_model_parallel_size` |\n| EP too low for large MoE | OOM on expert params | Increase EP; each expert lives on EP\u002Fnum_experts ranks |\n| CUDA graphs + packed sequences | Assert: \"CUDA graph accepts only Tensor inputs\" | Disable packing or use `local` full-iteration graphs |\n| CUDA graphs + full recompute | Assert: \"full recompute only with full iteration CUDA graph\" | Disable recompute or switch to `local` impl |\n| `use_te_rng_tracker` not set | Assert on provider init when CUDA graphs enabled | Set `cfg.model.use_te_rng_tracker = True` and `cfg.rng.te_rng_tracker = True` |\n| FSDP + TP > 1 on H100 | Possible comm bottleneck | Prefer FSDP with TP=1 or TP=2 on H100; FSDP shines on GB\u002FB-series |\n| Long context without CP | OOM on activations | Add CP=2\u002F4\u002F8; use `*_16k`, `*_64k`, or `*_128k` recipe variants |\n| MoE `overlap_grad_reduce` on H100 | May hurt perf (False in many H100 presets) | Set `overlap_grad_reduce=False` for MoE on H100 |\n| VLM SFT missing image data | Runs but produces garbage | Provide actual multimodal dataset or use mock VLM data |\n| Qwen35-VL MoE FSDP | Tested on Blackwell only | May not work on H100; validate first |\n\n### Recipe Override Examples\n\n```bash\n# Scale Llama3 8B from 2 GPUs to 8 GPUs (increase DP)\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe llama3_8b_pretrain_config \\\n    --dataset llm-pretrain-mock\n\n# Reduce parallelism for Qwen3-MoE 30B to fit on 4 GPUs\nuv run python -m torch.distributed.run --nproc_per_node=4 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe qwen3_30b_a3b_sft_config \\\n    --dataset llm-finetune \\\n    'model.expert_model_parallel_size=4'\n\n# Add long context to an existing recipe\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe llama3_8b_pretrain_config \\\n    --dataset llm-pretrain-mock \\\n    'model.seq_length=32768' \\\n    'model.context_parallel_size=4'\n\n# Enable CUDA graphs on any recipe\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe qwen3_30b_a3b_pretrain_config \\\n    --dataset llm-pretrain-mock \\\n    'model.cuda_graph_impl=transformer_engine' \\\n    'model.cuda_graph_scope=[attn,moe_router,moe_preprocess]' \\\n    'model.use_te_rng_tracker=True' \\\n    'rng.te_rng_tracker=True'\n```\n\n---\n\n## Quick Reference: Which Recipe for My Situation?\n\n| I want to... | Start with | GPUs needed |\n|---|---|---|\n| Try Bridge for the first time | `llama3_8b_sft_config` + mock data | 2 |\n| Fine-tune a 7-8B model | `llama3_8b_sft_config` or `qwen3_8b_sft_config` | 2–8 |\n| LoRA on 1 GPU | `llama3_8b_peft_config` or `qwen3_8b_peft_config` | 1 |\n| Pretrain a dense 70B | `llama3_70b_pretrain_config` | 32–64 |\n| Train a small MoE | `qwen3_30b_a3b_pretrain_config` | 8 |\n| Train a large MoE (235B+) | `qwen3_235b_a22b_pretrain_config` | 256–512 |\n| Benchmark throughput | Perf recipes via `run_script.py` | Varies |\n| Long-context training | `llama3_8b_128k_pretrain_config` or add CP override | 16+ |\n| VLM fine-tuning | `qwen3_vl_8b_sft_config` or `gemma3_vl_*_sft_config` | 4–8 |\n| Diffusion training | `wan_1_3B_pretrain_config` or `flux_12b_pretrain_config` | 8 |\n\n---\n\n## Code Anchors\n\n| What | Path |\n|------|------|\n| Library recipes root | `src\u002Fmegatron\u002Fbridge\u002Frecipes\u002F` |\n| Recipe `__init__.py` (all exports) | `src\u002Fmegatron\u002Fbridge\u002Frecipes\u002F__init__.py` |\n| Common recipe helpers | `src\u002Fmegatron\u002Fbridge\u002Frecipes\u002Fcommon.py` |\n| Training entry point | `scripts\u002Ftraining\u002Frun_recipe.py` |\n| Perf recipes root | `src\u002Fmegatron\u002Fbridge\u002Fperf_recipes\u002F` |\n| Perf entry point | `scripts\u002Fperformance\u002Frun_script.py` |\n| Perf recipe helpers | `scripts\u002Fperformance\u002Futils\u002Futils.py` |\n| Perf overrides (benchmark defaults) | `scripts\u002Fperformance\u002Futils\u002Foverrides.py` |\n",{"data":32,"body":34},{"name":4,"license":23,"description":6,"when_to_use":33},"User wants a starting recipe or training config; 'which recipe', 'recommend recipe', 'how to train Llama', 'starting config for X GPUs', 'what recipe for SFT'.",{"type":35,"children":36},"root",[37,46,52,59,112,118,123,239,243,249,256,598,604,705,710,713,719,731,811,821,826,829,835,855,861,1536,1542,1636,1642,1809,1815,2075,2081,2197,2203,2368,2374,2602,2608,2699,2705,2876,2882,3073,3079,3313,3319,3420,3423,3429,3449,3484,3490,3587,3592,3598,3648,3653,3747,3753,3823,3829,3878,3884,3976,3982,4032,4035,4041,4051,4054,4060,4066,4071,4173,4179,4225,4231,4534,4540,4998,5001,5007,5277,5280,5286,5448],{"type":38,"tag":39,"props":40,"children":42},"element","h1",{"id":41},"auto-recipe-recipe-index-recommendation",[43],{"type":44,"value":45},"text","Auto Recipe — Recipe Index & Recommendation",{"type":38,"tag":47,"props":48,"children":49},"p",{},[50],{"type":44,"value":51},"This skill indexes every shipped recipe and helps users pick the right starting\nconfig, adjust parallelism, and avoid common pitfalls.",{"type":38,"tag":53,"props":54,"children":56},"h2",{"id":55},"how-to-use-this-skill",[57],{"type":44,"value":58},"How to Use This Skill",{"type":38,"tag":60,"props":61,"children":62},"ol",{},[63,97,102,107],{"type":38,"tag":64,"props":65,"children":66},"li",{},[67,69,75,77,82,83,88,90,95],{"type":44,"value":68},"Ask the user for: ",{"type":38,"tag":70,"props":71,"children":72},"strong",{},[73],{"type":44,"value":74},"model name\u002Fsize",{"type":44,"value":76},", ",{"type":38,"tag":70,"props":78,"children":79},{},[80],{"type":44,"value":81},"GPU count & type",{"type":44,"value":76},{"type":38,"tag":70,"props":84,"children":85},{},[86],{"type":44,"value":87},"training goal",{"type":44,"value":89},"\n(pretrain \u002F SFT \u002F PEFT), and ",{"type":38,"tag":70,"props":91,"children":92},{},[93],{"type":44,"value":94},"sequence length",{"type":44,"value":96}," (if non-default).",{"type":38,"tag":64,"props":98,"children":99},{},[100],{"type":44,"value":101},"Look up the best-match recipe in the index below.",{"type":38,"tag":64,"props":103,"children":104},{},[105],{"type":44,"value":106},"Recommend the recipe function name + entry-point command.",{"type":38,"tag":64,"props":108,"children":109},{},[110],{"type":44,"value":111},"Provide adjustment advice (parallelism resizing, batch tuning, pitfalls).",{"type":38,"tag":53,"props":113,"children":115},{"id":114},"first-answer-checklist",[116],{"type":44,"value":117},"First Answer Checklist",{"type":38,"tag":47,"props":119,"children":120},{},[121],{"type":44,"value":122},"When recommending recipes, always include these distinctions before the long\nindex details:",{"type":38,"tag":60,"props":124,"children":125},{},[126,153,170,199,218],{"type":38,"tag":64,"props":127,"children":128},{},[129,134,136,143,145,151],{"type":38,"tag":70,"props":130,"children":131},{},[132],{"type":44,"value":133},"Library recipes",{"type":44,"value":135}," under ",{"type":38,"tag":137,"props":138,"children":140},"code",{"className":139},[],[141],{"type":44,"value":142},"src\u002Fmegatron\u002Fbridge\u002Frecipes\u002F",{"type":44,"value":144}," are for functional\ntraining and use ",{"type":38,"tag":137,"props":146,"children":148},{"className":147},[],[149],{"type":44,"value":150},"scripts\u002Ftraining\u002Frun_recipe.py",{"type":44,"value":152},".",{"type":38,"tag":64,"props":154,"children":155},{},[156,161,162,168],{"type":38,"tag":70,"props":157,"children":158},{},[159],{"type":44,"value":160},"Performance recipes",{"type":44,"value":135},{"type":38,"tag":137,"props":163,"children":165},{"className":164},[],[166],{"type":44,"value":167},"scripts\u002Fperformance\u002F",{"type":44,"value":169}," are for upper-bound\nthroughput benchmarks. They use mock data and should not be presented as\nproduction training recipes.",{"type":38,"tag":64,"props":171,"children":172},{},[173,175,181,183,189,191,197],{"type":44,"value":174},"For a first-time Bridge smoke test, recommend ",{"type":38,"tag":137,"props":176,"children":178},{"className":177},[],[179],{"type":44,"value":180},"llama3_8b_sft_config",{"type":44,"value":182}," with\nmock data via ",{"type":38,"tag":137,"props":184,"children":186},{"className":185},[],[187],{"type":44,"value":188},"--dataset llm-pretrain-mock",{"type":44,"value":190},". Do not use ",{"type":38,"tag":137,"props":192,"children":194},{"className":193},[],[195],{"type":44,"value":196},"llm-finetune",{"type":44,"value":198}," for\nthe setup-only tryout unless the user specifically asks for an SFT data path.",{"type":38,"tag":64,"props":200,"children":201},{},[202,204,210,212,217],{"type":44,"value":203},"For normal SFT recommendations, use ",{"type":38,"tag":137,"props":205,"children":207},{"className":206},[],[208],{"type":44,"value":209},"--dataset llm-finetune",{"type":44,"value":211},"; for pretrain\nand mock validation recommendations, use ",{"type":38,"tag":137,"props":213,"children":215},{"className":214},[],[216],{"type":44,"value":188},{"type":44,"value":152},{"type":38,"tag":64,"props":219,"children":220},{},[221,223,229,231,237],{"type":44,"value":222},"After the recipe and dataset, give the required resizing rules: TP must\ndivide ",{"type":38,"tag":137,"props":224,"children":226},{"className":225},[],[227],{"type":44,"value":228},"num_key_value_heads",{"type":44,"value":230},", keep TP within one node unless using\nNVL72-class interconnect, enable SP when TP > 1, configure CP for long\ncontext, DP is implicit, and reduce ",{"type":38,"tag":137,"props":232,"children":234},{"className":233},[],[235],{"type":44,"value":236},"micro_batch_size",{"type":44,"value":238}," first on OOM.",{"type":38,"tag":240,"props":241,"children":242},"hr",{},[],{"type":38,"tag":53,"props":244,"children":246},{"id":245},"entry-points",[247],{"type":44,"value":248},"Entry Points",{"type":38,"tag":250,"props":251,"children":253},"h3",{"id":252},"library-recipes-functional-training",[254],{"type":44,"value":255},"Library recipes (functional training)",{"type":38,"tag":257,"props":258,"children":263},"pre",{"className":259,"code":260,"language":261,"meta":262,"style":262},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","# Pretrain with mock data\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe \u003Crecipe_function_name> \\\n    --dataset llm-pretrain-mock\n\n# SFT with SQuAD\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe \u003Crecipe_function_name> \\\n    --dataset llm-finetune\n\n# Override any field via CLI\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe llama3_8b_pretrain_config \\\n    --dataset llm-pretrain-mock \\\n    'model.tensor_model_parallel_size=2' \\\n    'training.global_batch_size=64'\n","bash","",[264],{"type":38,"tag":137,"props":265,"children":266},{"__ignoreMap":262},[267,279,326,360,374,384,393,429,457,470,478,487,523,540,557,580],{"type":38,"tag":268,"props":269,"children":272},"span",{"class":270,"line":271},"line",1,[273],{"type":38,"tag":268,"props":274,"children":276},{"style":275},"--shiki-light:#90A4AE;--shiki-light-font-style:italic;--shiki-default:#546E7A;--shiki-default-font-style:italic;--shiki-dark:#676E95;--shiki-dark-font-style:italic",[277],{"type":44,"value":278},"# Pretrain with mock data\n",{"type":38,"tag":268,"props":280,"children":282},{"class":270,"line":281},2,[283,289,295,300,305,310,315,320],{"type":38,"tag":268,"props":284,"children":286},{"style":285},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[287],{"type":44,"value":288},"uv",{"type":38,"tag":268,"props":290,"children":292},{"style":291},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[293],{"type":44,"value":294}," run",{"type":38,"tag":268,"props":296,"children":297},{"style":291},[298],{"type":44,"value":299}," python",{"type":38,"tag":268,"props":301,"children":302},{"style":291},[303],{"type":44,"value":304}," -m",{"type":38,"tag":268,"props":306,"children":307},{"style":291},[308],{"type":44,"value":309}," torch.distributed.run",{"type":38,"tag":268,"props":311,"children":312},{"style":291},[313],{"type":44,"value":314}," --nproc_per_node=8",{"type":38,"tag":268,"props":316,"children":317},{"style":291},[318],{"type":44,"value":319}," scripts\u002Ftraining\u002Frun_recipe.py",{"type":38,"tag":268,"props":321,"children":323},{"style":322},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[324],{"type":44,"value":325}," \\\n",{"type":38,"tag":268,"props":327,"children":329},{"class":270,"line":328},3,[330,335,341,346,351,356],{"type":38,"tag":268,"props":331,"children":332},{"style":291},[333],{"type":44,"value":334},"    --recipe",{"type":38,"tag":268,"props":336,"children":338},{"style":337},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[339],{"type":44,"value":340}," \u003C",{"type":38,"tag":268,"props":342,"children":343},{"style":291},[344],{"type":44,"value":345},"recipe_function_nam",{"type":38,"tag":268,"props":347,"children":348},{"style":322},[349],{"type":44,"value":350},"e",{"type":38,"tag":268,"props":352,"children":353},{"style":337},[354],{"type":44,"value":355},">",{"type":38,"tag":268,"props":357,"children":358},{"style":322},[359],{"type":44,"value":325},{"type":38,"tag":268,"props":361,"children":363},{"class":270,"line":362},4,[364,369],{"type":38,"tag":268,"props":365,"children":366},{"style":291},[367],{"type":44,"value":368},"    --dataset",{"type":38,"tag":268,"props":370,"children":371},{"style":291},[372],{"type":44,"value":373}," llm-pretrain-mock\n",{"type":38,"tag":268,"props":375,"children":377},{"class":270,"line":376},5,[378],{"type":38,"tag":268,"props":379,"children":381},{"emptyLinePlaceholder":380},true,[382],{"type":44,"value":383},"\n",{"type":38,"tag":268,"props":385,"children":387},{"class":270,"line":386},6,[388],{"type":38,"tag":268,"props":389,"children":390},{"style":275},[391],{"type":44,"value":392},"# SFT with SQuAD\n",{"type":38,"tag":268,"props":394,"children":396},{"class":270,"line":395},7,[397,401,405,409,413,417,421,425],{"type":38,"tag":268,"props":398,"children":399},{"style":285},[400],{"type":44,"value":288},{"type":38,"tag":268,"props":402,"children":403},{"style":291},[404],{"type":44,"value":294},{"type":38,"tag":268,"props":406,"children":407},{"style":291},[408],{"type":44,"value":299},{"type":38,"tag":268,"props":410,"children":411},{"style":291},[412],{"type":44,"value":304},{"type":38,"tag":268,"props":414,"children":415},{"style":291},[416],{"type":44,"value":309},{"type":38,"tag":268,"props":418,"children":419},{"style":291},[420],{"type":44,"value":314},{"type":38,"tag":268,"props":422,"children":423},{"style":291},[424],{"type":44,"value":319},{"type":38,"tag":268,"props":426,"children":427},{"style":322},[428],{"type":44,"value":325},{"type":38,"tag":268,"props":430,"children":432},{"class":270,"line":431},8,[433,437,441,445,449,453],{"type":38,"tag":268,"props":434,"children":435},{"style":291},[436],{"type":44,"value":334},{"type":38,"tag":268,"props":438,"children":439},{"style":337},[440],{"type":44,"value":340},{"type":38,"tag":268,"props":442,"children":443},{"style":291},[444],{"type":44,"value":345},{"type":38,"tag":268,"props":446,"children":447},{"style":322},[448],{"type":44,"value":350},{"type":38,"tag":268,"props":450,"children":451},{"style":337},[452],{"type":44,"value":355},{"type":38,"tag":268,"props":454,"children":455},{"style":322},[456],{"type":44,"value":325},{"type":38,"tag":268,"props":458,"children":460},{"class":270,"line":459},9,[461,465],{"type":38,"tag":268,"props":462,"children":463},{"style":291},[464],{"type":44,"value":368},{"type":38,"tag":268,"props":466,"children":467},{"style":291},[468],{"type":44,"value":469}," llm-finetune\n",{"type":38,"tag":268,"props":471,"children":473},{"class":270,"line":472},10,[474],{"type":38,"tag":268,"props":475,"children":476},{"emptyLinePlaceholder":380},[477],{"type":44,"value":383},{"type":38,"tag":268,"props":479,"children":481},{"class":270,"line":480},11,[482],{"type":38,"tag":268,"props":483,"children":484},{"style":275},[485],{"type":44,"value":486},"# Override any field via CLI\n",{"type":38,"tag":268,"props":488,"children":490},{"class":270,"line":489},12,[491,495,499,503,507,511,515,519],{"type":38,"tag":268,"props":492,"children":493},{"style":285},[494],{"type":44,"value":288},{"type":38,"tag":268,"props":496,"children":497},{"style":291},[498],{"type":44,"value":294},{"type":38,"tag":268,"props":500,"children":501},{"style":291},[502],{"type":44,"value":299},{"type":38,"tag":268,"props":504,"children":505},{"style":291},[506],{"type":44,"value":304},{"type":38,"tag":268,"props":508,"children":509},{"style":291},[510],{"type":44,"value":309},{"type":38,"tag":268,"props":512,"children":513},{"style":291},[514],{"type":44,"value":314},{"type":38,"tag":268,"props":516,"children":517},{"style":291},[518],{"type":44,"value":319},{"type":38,"tag":268,"props":520,"children":521},{"style":322},[522],{"type":44,"value":325},{"type":38,"tag":268,"props":524,"children":526},{"class":270,"line":525},13,[527,531,536],{"type":38,"tag":268,"props":528,"children":529},{"style":291},[530],{"type":44,"value":334},{"type":38,"tag":268,"props":532,"children":533},{"style":291},[534],{"type":44,"value":535}," llama3_8b_pretrain_config",{"type":38,"tag":268,"props":537,"children":538},{"style":322},[539],{"type":44,"value":325},{"type":38,"tag":268,"props":541,"children":543},{"class":270,"line":542},14,[544,548,553],{"type":38,"tag":268,"props":545,"children":546},{"style":291},[547],{"type":44,"value":368},{"type":38,"tag":268,"props":549,"children":550},{"style":291},[551],{"type":44,"value":552}," llm-pretrain-mock",{"type":38,"tag":268,"props":554,"children":555},{"style":322},[556],{"type":44,"value":325},{"type":38,"tag":268,"props":558,"children":560},{"class":270,"line":559},15,[561,566,571,576],{"type":38,"tag":268,"props":562,"children":563},{"style":337},[564],{"type":44,"value":565},"    '",{"type":38,"tag":268,"props":567,"children":568},{"style":291},[569],{"type":44,"value":570},"model.tensor_model_parallel_size=2",{"type":38,"tag":268,"props":572,"children":573},{"style":337},[574],{"type":44,"value":575},"'",{"type":38,"tag":268,"props":577,"children":578},{"style":322},[579],{"type":44,"value":325},{"type":38,"tag":268,"props":581,"children":583},{"class":270,"line":582},16,[584,588,593],{"type":38,"tag":268,"props":585,"children":586},{"style":337},[587],{"type":44,"value":565},{"type":38,"tag":268,"props":589,"children":590},{"style":291},[591],{"type":44,"value":592},"training.global_batch_size=64",{"type":38,"tag":268,"props":594,"children":595},{"style":337},[596],{"type":44,"value":597},"'\n",{"type":38,"tag":250,"props":599,"children":601},{"id":600},"performance-recipes-throughput-benchmarks",[602],{"type":44,"value":603},"Performance recipes (throughput benchmarks)",{"type":38,"tag":257,"props":605,"children":607},{"className":259,"code":606,"language":261,"meta":262,"style":262},"python scripts\u002Fperformance\u002Frun_script.py \\\n    --recipe \u003Cmodel_family> \\\n    --gpu_type h100 \\\n    --num_gpus 64 \\\n    --data mock\n",[608],{"type":38,"tag":137,"props":609,"children":610},{"__ignoreMap":262},[611,628,657,674,692],{"type":38,"tag":268,"props":612,"children":613},{"class":270,"line":271},[614,619,624],{"type":38,"tag":268,"props":615,"children":616},{"style":285},[617],{"type":44,"value":618},"python",{"type":38,"tag":268,"props":620,"children":621},{"style":291},[622],{"type":44,"value":623}," scripts\u002Fperformance\u002Frun_script.py",{"type":38,"tag":268,"props":625,"children":626},{"style":322},[627],{"type":44,"value":325},{"type":38,"tag":268,"props":629,"children":630},{"class":270,"line":281},[631,635,639,644,649,653],{"type":38,"tag":268,"props":632,"children":633},{"style":291},[634],{"type":44,"value":334},{"type":38,"tag":268,"props":636,"children":637},{"style":337},[638],{"type":44,"value":340},{"type":38,"tag":268,"props":640,"children":641},{"style":291},[642],{"type":44,"value":643},"model_famil",{"type":38,"tag":268,"props":645,"children":646},{"style":322},[647],{"type":44,"value":648},"y",{"type":38,"tag":268,"props":650,"children":651},{"style":337},[652],{"type":44,"value":355},{"type":38,"tag":268,"props":654,"children":655},{"style":322},[656],{"type":44,"value":325},{"type":38,"tag":268,"props":658,"children":659},{"class":270,"line":328},[660,665,670],{"type":38,"tag":268,"props":661,"children":662},{"style":291},[663],{"type":44,"value":664},"    --gpu_type",{"type":38,"tag":268,"props":666,"children":667},{"style":291},[668],{"type":44,"value":669}," h100",{"type":38,"tag":268,"props":671,"children":672},{"style":322},[673],{"type":44,"value":325},{"type":38,"tag":268,"props":675,"children":676},{"class":270,"line":362},[677,682,688],{"type":38,"tag":268,"props":678,"children":679},{"style":291},[680],{"type":44,"value":681},"    --num_gpus",{"type":38,"tag":268,"props":683,"children":685},{"style":684},"--shiki-light:#F76D47;--shiki-default:#F78C6C;--shiki-dark:#F78C6C",[686],{"type":44,"value":687}," 64",{"type":38,"tag":268,"props":689,"children":690},{"style":322},[691],{"type":44,"value":325},{"type":38,"tag":268,"props":693,"children":694},{"class":270,"line":376},[695,700],{"type":38,"tag":268,"props":696,"children":697},{"style":291},[698],{"type":44,"value":699},"    --data",{"type":38,"tag":268,"props":701,"children":702},{"style":291},[703],{"type":44,"value":704}," mock\n",{"type":38,"tag":47,"props":706,"children":707},{},[708],{"type":44,"value":709},"See the Performance Recipe Index for important caveats before using these for anything beyond throughput benchmarking.",{"type":38,"tag":240,"props":711,"children":712},{},[],{"type":38,"tag":53,"props":714,"children":716},{"id":715},"perf-recipe-layout",[717],{"type":44,"value":718},"Perf Recipe Layout",{"type":38,"tag":47,"props":720,"children":721},{},[722,724,729],{"type":44,"value":723},"Performance recipes use the same ",{"type":38,"tag":70,"props":725,"children":726},{},[727],{"type":44,"value":728},"Python function",{"type":44,"value":730}," format as library recipes,\nbut live in a dedicated namespace for throughput benchmarking:",{"type":38,"tag":732,"props":733,"children":734},"ul",{},[735,746,766,771,790],{"type":38,"tag":64,"props":736,"children":737},{},[738,740],{"type":44,"value":739},"Perf recipes live in ",{"type":38,"tag":137,"props":741,"children":743},{"className":742},[],[744],{"type":44,"value":745},"src\u002Fmegatron\u002Fbridge\u002Fperf_recipes\u002F\u003Cfamily>\u002F\u003Chardware>\u002F\u003Cmodel>.py",{"type":38,"tag":64,"props":747,"children":748},{},[749,751,756,758,764],{"type":44,"value":750},"Each perf recipe is a ",{"type":38,"tag":70,"props":752,"children":753},{},[754],{"type":44,"value":755},"self-contained Python function",{"type":44,"value":757}," (e.g. ",{"type":38,"tag":137,"props":759,"children":761},{"className":760},[],[762],{"type":44,"value":763},"llama3_8b_pretrain_8gpu_h100_bf16_config()",{"type":44,"value":765},")",{"type":38,"tag":64,"props":767,"children":768},{},[769],{"type":44,"value":770},"Recipe names encode model, task, GPU count, hardware, precision, and optional variant",{"type":38,"tag":64,"props":772,"children":773},{},[774,780,782,788],{"type":38,"tag":137,"props":775,"children":777},{"className":776},[],[778],{"type":44,"value":779},"scripts\u002Fperformance\u002Futils\u002Futils.py",{"type":44,"value":781}," derives compatibility ",{"type":38,"tag":137,"props":783,"children":785},{"className":784},[],[786],{"type":44,"value":787},"WorkloadBaseConfig",{"type":44,"value":789}," views from the flat recipe itself",{"type":38,"tag":64,"props":791,"children":792},{},[793,795,801,803,809],{"type":44,"value":794},"Shared helpers: ",{"type":38,"tag":137,"props":796,"children":798},{"className":797},[],[799],{"type":44,"value":800},"_benchmark_common()",{"type":44,"value":802}," (50 iters, timing, TE RNG), ",{"type":38,"tag":137,"props":804,"children":806},{"className":805},[],[807],{"type":44,"value":808},"_perf_precision()",{"type":44,"value":810}," (bf16 \u002F fp8_cs \u002F fp8_mx \u002F nvfp4)",{"type":38,"tag":47,"props":812,"children":813},{},[814,819],{"type":38,"tag":70,"props":815,"children":816},{},[817],{"type":44,"value":818},"Why Python, not YAML?",{"type":44,"value":820}," Previous YAML-based approaches had problems:\nrecipe logic was split across multiple indirection layers, configs were not\nself-contained, and the two-level pipeline made maintenance and debugging\ndifficult. Python functions are explicit, greppable, and composable.",{"type":38,"tag":47,"props":822,"children":823},{},[824],{"type":44,"value":825},"The training launcher can invoke both library recipes and perf recipes without\nthe removed legacy config package.",{"type":38,"tag":240,"props":827,"children":828},{},[],{"type":38,"tag":53,"props":830,"children":832},{"id":831},"library-recipe-index",[833],{"type":44,"value":834},"Library Recipe Index",{"type":38,"tag":47,"props":836,"children":837},{},[838,840,845,847,853],{"type":44,"value":839},"All recipes live under ",{"type":38,"tag":137,"props":841,"children":843},{"className":842},[],[844],{"type":44,"value":142},{"type":44,"value":846},". Each function returns a\n",{"type":38,"tag":137,"props":848,"children":850},{"className":849},[],[851],{"type":44,"value":852},"ConfigContainer",{"type":44,"value":854}," with model, training, optimizer, and data settings.",{"type":38,"tag":250,"props":856,"children":858},{"id":857},"llama",[859],{"type":44,"value":860},"Llama",{"type":38,"tag":862,"props":863,"children":864},"table",{},[865,914],{"type":38,"tag":866,"props":867,"children":868},"thead",{},[869],{"type":38,"tag":870,"props":871,"children":872},"tr",{},[873,879,884,889,894,899,904,909],{"type":38,"tag":874,"props":875,"children":876},"th",{},[877],{"type":44,"value":878},"Recipe",{"type":38,"tag":874,"props":880,"children":881},{},[882],{"type":44,"value":883},"Mode",{"type":38,"tag":874,"props":885,"children":886},{},[887],{"type":44,"value":888},"TP",{"type":38,"tag":874,"props":890,"children":891},{},[892],{"type":44,"value":893},"PP",{"type":38,"tag":874,"props":895,"children":896},{},[897],{"type":44,"value":898},"CP",{"type":38,"tag":874,"props":900,"children":901},{},[902],{"type":44,"value":903},"SP",{"type":38,"tag":874,"props":905,"children":906},{},[907],{"type":44,"value":908},"GPUs (min)",{"type":38,"tag":874,"props":910,"children":911},{},[912],{"type":44,"value":913},"Seq Len",{"type":38,"tag":915,"props":916,"children":917},"tbody",{},[918,964,1006,1048,1090,1132,1173,1214,1255,1295,1335,1375,1415,1456,1496],{"type":38,"tag":870,"props":919,"children":920},{},[921,931,936,941,946,951,955,959],{"type":38,"tag":922,"props":923,"children":924},"td",{},[925],{"type":38,"tag":137,"props":926,"children":928},{"className":927},[],[929],{"type":44,"value":930},"llama2_7b_pretrain_config",{"type":38,"tag":922,"props":932,"children":933},{},[934],{"type":44,"value":935},"Pretrain",{"type":38,"tag":922,"props":937,"children":938},{},[939],{"type":44,"value":940},"2",{"type":38,"tag":922,"props":942,"children":943},{},[944],{"type":44,"value":945},"1",{"type":38,"tag":922,"props":947,"children":948},{},[949],{"type":44,"value":950},"—",{"type":38,"tag":922,"props":952,"children":953},{},[954],{"type":44,"value":950},{"type":38,"tag":922,"props":956,"children":957},{},[958],{"type":44,"value":940},{"type":38,"tag":922,"props":960,"children":961},{},[962],{"type":44,"value":963},"4K",{"type":38,"tag":870,"props":965,"children":966},{},[967,976,980,984,988,992,997,1001],{"type":38,"tag":922,"props":968,"children":969},{},[970],{"type":38,"tag":137,"props":971,"children":973},{"className":972},[],[974],{"type":44,"value":975},"llama3_8b_pretrain_config",{"type":38,"tag":922,"props":977,"children":978},{},[979],{"type":44,"value":935},{"type":38,"tag":922,"props":981,"children":982},{},[983],{"type":44,"value":940},{"type":38,"tag":922,"props":985,"children":986},{},[987],{"type":44,"value":945},{"type":38,"tag":922,"props":989,"children":990},{},[991],{"type":44,"value":950},{"type":38,"tag":922,"props":993,"children":994},{},[995],{"type":44,"value":996},"✓",{"type":38,"tag":922,"props":998,"children":999},{},[1000],{"type":44,"value":940},{"type":38,"tag":922,"props":1002,"children":1003},{},[1004],{"type":44,"value":1005},"8K",{"type":38,"tag":870,"props":1007,"children":1008},{},[1009,1018,1022,1026,1030,1034,1038,1043],{"type":38,"tag":922,"props":1010,"children":1011},{},[1012],{"type":38,"tag":137,"props":1013,"children":1015},{"className":1014},[],[1016],{"type":44,"value":1017},"llama3_8b_16k_pretrain_config",{"type":38,"tag":922,"props":1019,"children":1020},{},[1021],{"type":44,"value":935},{"type":38,"tag":922,"props":1023,"children":1024},{},[1025],{"type":44,"value":940},{"type":38,"tag":922,"props":1027,"children":1028},{},[1029],{"type":44,"value":945},{"type":38,"tag":922,"props":1031,"children":1032},{},[1033],{"type":44,"value":940},{"type":38,"tag":922,"props":1035,"children":1036},{},[1037],{"type":44,"value":996},{"type":38,"tag":922,"props":1039,"children":1040},{},[1041],{"type":44,"value":1042},"4",{"type":38,"tag":922,"props":1044,"children":1045},{},[1046],{"type":44,"value":1047},"16K",{"type":38,"tag":870,"props":1049,"children":1050},{},[1051,1060,1064,1068,1072,1076,1080,1085],{"type":38,"tag":922,"props":1052,"children":1053},{},[1054],{"type":38,"tag":137,"props":1055,"children":1057},{"className":1056},[],[1058],{"type":44,"value":1059},"llama3_8b_64k_pretrain_config",{"type":38,"tag":922,"props":1061,"children":1062},{},[1063],{"type":44,"value":935},{"type":38,"tag":922,"props":1065,"children":1066},{},[1067],{"type":44,"value":940},{"type":38,"tag":922,"props":1069,"children":1070},{},[1071],{"type":44,"value":945},{"type":38,"tag":922,"props":1073,"children":1074},{},[1075],{"type":44,"value":1042},{"type":38,"tag":922,"props":1077,"children":1078},{},[1079],{"type":44,"value":996},{"type":38,"tag":922,"props":1081,"children":1082},{},[1083],{"type":44,"value":1084},"8",{"type":38,"tag":922,"props":1086,"children":1087},{},[1088],{"type":44,"value":1089},"64K",{"type":38,"tag":870,"props":1091,"children":1092},{},[1093,1102,1106,1110,1114,1118,1122,1127],{"type":38,"tag":922,"props":1094,"children":1095},{},[1096],{"type":38,"tag":137,"props":1097,"children":1099},{"className":1098},[],[1100],{"type":44,"value":1101},"llama3_8b_128k_pretrain_config",{"type":38,"tag":922,"props":1103,"children":1104},{},[1105],{"type":44,"value":935},{"type":38,"tag":922,"props":1107,"children":1108},{},[1109],{"type":44,"value":940},{"type":38,"tag":922,"props":1111,"children":1112},{},[1113],{"type":44,"value":945},{"type":38,"tag":922,"props":1115,"children":1116},{},[1117],{"type":44,"value":1084},{"type":38,"tag":922,"props":1119,"children":1120},{},[1121],{"type":44,"value":996},{"type":38,"tag":922,"props":1123,"children":1124},{},[1125],{"type":44,"value":1126},"16",{"type":38,"tag":922,"props":1128,"children":1129},{},[1130],{"type":44,"value":1131},"128K",{"type":38,"tag":870,"props":1133,"children":1134},{},[1135,1144,1148,1152,1156,1160,1164,1169],{"type":38,"tag":922,"props":1136,"children":1137},{},[1138],{"type":38,"tag":137,"props":1139,"children":1141},{"className":1140},[],[1142],{"type":44,"value":1143},"llama3_70b_pretrain_config",{"type":38,"tag":922,"props":1145,"children":1146},{},[1147],{"type":44,"value":935},{"type":38,"tag":922,"props":1149,"children":1150},{},[1151],{"type":44,"value":1084},{"type":38,"tag":922,"props":1153,"children":1154},{},[1155],{"type":44,"value":1042},{"type":38,"tag":922,"props":1157,"children":1158},{},[1159],{"type":44,"value":950},{"type":38,"tag":922,"props":1161,"children":1162},{},[1163],{"type":44,"value":996},{"type":38,"tag":922,"props":1165,"children":1166},{},[1167],{"type":44,"value":1168},"32",{"type":38,"tag":922,"props":1170,"children":1171},{},[1172],{"type":44,"value":1005},{"type":38,"tag":870,"props":1174,"children":1175},{},[1176,1185,1189,1193,1197,1201,1205,1210],{"type":38,"tag":922,"props":1177,"children":1178},{},[1179],{"type":38,"tag":137,"props":1180,"children":1182},{"className":1181},[],[1183],{"type":44,"value":1184},"llama3_70b_16k_pretrain_config",{"type":38,"tag":922,"props":1186,"children":1187},{},[1188],{"type":44,"value":935},{"type":38,"tag":922,"props":1190,"children":1191},{},[1192],{"type":44,"value":1084},{"type":38,"tag":922,"props":1194,"children":1195},{},[1196],{"type":44,"value":1042},{"type":38,"tag":922,"props":1198,"children":1199},{},[1200],{"type":44,"value":940},{"type":38,"tag":922,"props":1202,"children":1203},{},[1204],{"type":44,"value":996},{"type":38,"tag":922,"props":1206,"children":1207},{},[1208],{"type":44,"value":1209},"64",{"type":38,"tag":922,"props":1211,"children":1212},{},[1213],{"type":44,"value":1047},{"type":38,"tag":870,"props":1215,"children":1216},{},[1217,1226,1230,1234,1238,1242,1246,1251],{"type":38,"tag":922,"props":1218,"children":1219},{},[1220],{"type":38,"tag":137,"props":1221,"children":1223},{"className":1222},[],[1224],{"type":44,"value":1225},"llama3_70b_64k_pretrain_config",{"type":38,"tag":922,"props":1227,"children":1228},{},[1229],{"type":44,"value":935},{"type":38,"tag":922,"props":1231,"children":1232},{},[1233],{"type":44,"value":1084},{"type":38,"tag":922,"props":1235,"children":1236},{},[1237],{"type":44,"value":1042},{"type":38,"tag":922,"props":1239,"children":1240},{},[1241],{"type":44,"value":1042},{"type":38,"tag":922,"props":1243,"children":1244},{},[1245],{"type":44,"value":996},{"type":38,"tag":922,"props":1247,"children":1248},{},[1249],{"type":44,"value":1250},"128",{"type":38,"tag":922,"props":1252,"children":1253},{},[1254],{"type":44,"value":1089},{"type":38,"tag":870,"props":1256,"children":1257},{},[1258,1267,1271,1275,1279,1283,1287,1291],{"type":38,"tag":922,"props":1259,"children":1260},{},[1261],{"type":38,"tag":137,"props":1262,"children":1264},{"className":1263},[],[1265],{"type":44,"value":1266},"llama31_405b_pretrain_config",{"type":38,"tag":922,"props":1268,"children":1269},{},[1270],{"type":44,"value":935},{"type":38,"tag":922,"props":1272,"children":1273},{},[1274],{"type":44,"value":1084},{"type":38,"tag":922,"props":1276,"children":1277},{},[1278],{"type":44,"value":1126},{"type":38,"tag":922,"props":1280,"children":1281},{},[1282],{"type":44,"value":950},{"type":38,"tag":922,"props":1284,"children":1285},{},[1286],{"type":44,"value":996},{"type":38,"tag":922,"props":1288,"children":1289},{},[1290],{"type":44,"value":1250},{"type":38,"tag":922,"props":1292,"children":1293},{},[1294],{"type":44,"value":1005},{"type":38,"tag":870,"props":1296,"children":1297},{},[1298,1306,1311,1315,1319,1323,1327,1331],{"type":38,"tag":922,"props":1299,"children":1300},{},[1301],{"type":38,"tag":137,"props":1302,"children":1304},{"className":1303},[],[1305],{"type":44,"value":180},{"type":38,"tag":922,"props":1307,"children":1308},{},[1309],{"type":44,"value":1310},"SFT",{"type":38,"tag":922,"props":1312,"children":1313},{},[1314],{"type":44,"value":940},{"type":38,"tag":922,"props":1316,"children":1317},{},[1318],{"type":44,"value":945},{"type":38,"tag":922,"props":1320,"children":1321},{},[1322],{"type":44,"value":950},{"type":38,"tag":922,"props":1324,"children":1325},{},[1326],{"type":44,"value":996},{"type":38,"tag":922,"props":1328,"children":1329},{},[1330],{"type":44,"value":940},{"type":38,"tag":922,"props":1332,"children":1333},{},[1334],{"type":44,"value":1005},{"type":38,"tag":870,"props":1336,"children":1337},{},[1338,1347,1351,1355,1359,1363,1367,1371],{"type":38,"tag":922,"props":1339,"children":1340},{},[1341],{"type":38,"tag":137,"props":1342,"children":1344},{"className":1343},[],[1345],{"type":44,"value":1346},"llama3_70b_sft_config",{"type":38,"tag":922,"props":1348,"children":1349},{},[1350],{"type":44,"value":1310},{"type":38,"tag":922,"props":1352,"children":1353},{},[1354],{"type":44,"value":1042},{"type":38,"tag":922,"props":1356,"children":1357},{},[1358],{"type":44,"value":1042},{"type":38,"tag":922,"props":1360,"children":1361},{},[1362],{"type":44,"value":950},{"type":38,"tag":922,"props":1364,"children":1365},{},[1366],{"type":44,"value":996},{"type":38,"tag":922,"props":1368,"children":1369},{},[1370],{"type":44,"value":1126},{"type":38,"tag":922,"props":1372,"children":1373},{},[1374],{"type":44,"value":1005},{"type":38,"tag":870,"props":1376,"children":1377},{},[1378,1387,1391,1395,1399,1403,1407,1411],{"type":38,"tag":922,"props":1379,"children":1380},{},[1381],{"type":38,"tag":137,"props":1382,"children":1384},{"className":1383},[],[1385],{"type":44,"value":1386},"llama31_405b_sft_config",{"type":38,"tag":922,"props":1388,"children":1389},{},[1390],{"type":44,"value":1310},{"type":38,"tag":922,"props":1392,"children":1393},{},[1394],{"type":44,"value":1084},{"type":38,"tag":922,"props":1396,"children":1397},{},[1398],{"type":44,"value":1084},{"type":38,"tag":922,"props":1400,"children":1401},{},[1402],{"type":44,"value":950},{"type":38,"tag":922,"props":1404,"children":1405},{},[1406],{"type":44,"value":996},{"type":38,"tag":922,"props":1408,"children":1409},{},[1410],{"type":44,"value":1209},{"type":38,"tag":922,"props":1412,"children":1413},{},[1414],{"type":44,"value":1005},{"type":38,"tag":870,"props":1416,"children":1417},{},[1418,1427,1432,1436,1440,1444,1448,1452],{"type":38,"tag":922,"props":1419,"children":1420},{},[1421],{"type":38,"tag":137,"props":1422,"children":1424},{"className":1423},[],[1425],{"type":44,"value":1426},"llama3_8b_peft_config",{"type":38,"tag":922,"props":1428,"children":1429},{},[1430],{"type":44,"value":1431},"PEFT",{"type":38,"tag":922,"props":1433,"children":1434},{},[1435],{"type":44,"value":945},{"type":38,"tag":922,"props":1437,"children":1438},{},[1439],{"type":44,"value":945},{"type":38,"tag":922,"props":1441,"children":1442},{},[1443],{"type":44,"value":950},{"type":38,"tag":922,"props":1445,"children":1446},{},[1447],{"type":44,"value":950},{"type":38,"tag":922,"props":1449,"children":1450},{},[1451],{"type":44,"value":945},{"type":38,"tag":922,"props":1453,"children":1454},{},[1455],{"type":44,"value":1005},{"type":38,"tag":870,"props":1457,"children":1458},{},[1459,1468,1472,1476,1480,1484,1488,1492],{"type":38,"tag":922,"props":1460,"children":1461},{},[1462],{"type":38,"tag":137,"props":1463,"children":1465},{"className":1464},[],[1466],{"type":44,"value":1467},"llama3_70b_peft_config",{"type":38,"tag":922,"props":1469,"children":1470},{},[1471],{"type":44,"value":1431},{"type":38,"tag":922,"props":1473,"children":1474},{},[1475],{"type":44,"value":940},{"type":38,"tag":922,"props":1477,"children":1478},{},[1479],{"type":44,"value":1042},{"type":38,"tag":922,"props":1481,"children":1482},{},[1483],{"type":44,"value":950},{"type":38,"tag":922,"props":1485,"children":1486},{},[1487],{"type":44,"value":996},{"type":38,"tag":922,"props":1489,"children":1490},{},[1491],{"type":44,"value":1084},{"type":38,"tag":922,"props":1493,"children":1494},{},[1495],{"type":44,"value":1005},{"type":38,"tag":870,"props":1497,"children":1498},{},[1499,1508,1512,1516,1520,1524,1528,1532],{"type":38,"tag":922,"props":1500,"children":1501},{},[1502],{"type":38,"tag":137,"props":1503,"children":1505},{"className":1504},[],[1506],{"type":44,"value":1507},"llama31_405b_peft_config",{"type":38,"tag":922,"props":1509,"children":1510},{},[1511],{"type":44,"value":1431},{"type":38,"tag":922,"props":1513,"children":1514},{},[1515],{"type":44,"value":1042},{"type":38,"tag":922,"props":1517,"children":1518},{},[1519],{"type":44,"value":1084},{"type":38,"tag":922,"props":1521,"children":1522},{},[1523],{"type":44,"value":950},{"type":38,"tag":922,"props":1525,"children":1526},{},[1527],{"type":44,"value":996},{"type":38,"tag":922,"props":1529,"children":1530},{},[1531],{"type":44,"value":1168},{"type":38,"tag":922,"props":1533,"children":1534},{},[1535],{"type":44,"value":1005},{"type":38,"tag":250,"props":1537,"children":1539},{"id":1538},"qwen2-qwen25",[1540],{"type":44,"value":1541},"Qwen2 \u002F Qwen2.5",{"type":38,"tag":862,"props":1543,"children":1544},{},[1545,1572],{"type":38,"tag":866,"props":1546,"children":1547},{},[1548],{"type":38,"tag":870,"props":1549,"children":1550},{},[1551,1555,1559,1563,1567],{"type":38,"tag":874,"props":1552,"children":1553},{},[1554],{"type":44,"value":878},{"type":38,"tag":874,"props":1556,"children":1557},{},[1558],{"type":44,"value":883},{"type":38,"tag":874,"props":1560,"children":1561},{},[1562],{"type":44,"value":888},{"type":38,"tag":874,"props":1564,"children":1565},{},[1566],{"type":44,"value":893},{"type":38,"tag":874,"props":1568,"children":1569},{},[1570],{"type":44,"value":1571},"Sizes",{"type":38,"tag":915,"props":1573,"children":1574},{},[1575,1607],{"type":38,"tag":870,"props":1576,"children":1577},{},[1578,1587,1592,1597,1602],{"type":38,"tag":922,"props":1579,"children":1580},{},[1581],{"type":38,"tag":137,"props":1582,"children":1584},{"className":1583},[],[1585],{"type":44,"value":1586},"qwen2_*_{pretrain,sft,peft}_config",{"type":38,"tag":922,"props":1588,"children":1589},{},[1590],{"type":44,"value":1591},"All",{"type":38,"tag":922,"props":1593,"children":1594},{},[1595],{"type":44,"value":1596},"1–8",{"type":38,"tag":922,"props":1598,"children":1599},{},[1600],{"type":44,"value":1601},"1–4",{"type":38,"tag":922,"props":1603,"children":1604},{},[1605],{"type":44,"value":1606},"500M, 1.5B, 7B, 14B, 32B, 72B",{"type":38,"tag":870,"props":1608,"children":1609},{},[1610,1619,1623,1627,1631],{"type":38,"tag":922,"props":1611,"children":1612},{},[1613],{"type":38,"tag":137,"props":1614,"children":1616},{"className":1615},[],[1617],{"type":44,"value":1618},"qwen25_*_{pretrain,sft,peft}_config",{"type":38,"tag":922,"props":1620,"children":1621},{},[1622],{"type":44,"value":1591},{"type":38,"tag":922,"props":1624,"children":1625},{},[1626],{"type":44,"value":1596},{"type":38,"tag":922,"props":1628,"children":1629},{},[1630],{"type":44,"value":1601},{"type":38,"tag":922,"props":1632,"children":1633},{},[1634],{"type":44,"value":1635},"500M, 1.5B, 3B, 7B, 14B, 32B, 72B",{"type":38,"tag":250,"props":1637,"children":1639},{"id":1638},"qwen3-dense",[1640],{"type":44,"value":1641},"Qwen3 (Dense)",{"type":38,"tag":862,"props":1643,"children":1644},{},[1645,1675],{"type":38,"tag":866,"props":1646,"children":1647},{},[1648],{"type":38,"tag":870,"props":1649,"children":1650},{},[1651,1655,1659,1663,1667,1671],{"type":38,"tag":874,"props":1652,"children":1653},{},[1654],{"type":44,"value":878},{"type":38,"tag":874,"props":1656,"children":1657},{},[1658],{"type":44,"value":883},{"type":38,"tag":874,"props":1660,"children":1661},{},[1662],{"type":44,"value":888},{"type":38,"tag":874,"props":1664,"children":1665},{},[1666],{"type":44,"value":893},{"type":38,"tag":874,"props":1668,"children":1669},{},[1670],{"type":44,"value":898},{"type":38,"tag":874,"props":1672,"children":1673},{},[1674],{"type":44,"value":1571},{"type":38,"tag":915,"props":1676,"children":1677},{},[1678,1712,1744,1777],{"type":38,"tag":870,"props":1679,"children":1680},{},[1681,1690,1694,1698,1703,1707],{"type":38,"tag":922,"props":1682,"children":1683},{},[1684],{"type":38,"tag":137,"props":1685,"children":1687},{"className":1686},[],[1688],{"type":44,"value":1689},"qwen3_*_pretrain_config",{"type":38,"tag":922,"props":1691,"children":1692},{},[1693],{"type":44,"value":935},{"type":38,"tag":922,"props":1695,"children":1696},{},[1697],{"type":44,"value":1596},{"type":38,"tag":922,"props":1699,"children":1700},{},[1701],{"type":44,"value":1702},"1–2",{"type":38,"tag":922,"props":1704,"children":1705},{},[1706],{"type":44,"value":950},{"type":38,"tag":922,"props":1708,"children":1709},{},[1710],{"type":44,"value":1711},"600M–32B",{"type":38,"tag":870,"props":1713,"children":1714},{},[1715,1724,1728,1732,1736,1740],{"type":38,"tag":922,"props":1716,"children":1717},{},[1718],{"type":38,"tag":137,"props":1719,"children":1721},{"className":1720},[],[1722],{"type":44,"value":1723},"qwen3_*_sft_config",{"type":38,"tag":922,"props":1725,"children":1726},{},[1727],{"type":44,"value":1310},{"type":38,"tag":922,"props":1729,"children":1730},{},[1731],{"type":44,"value":1596},{"type":38,"tag":922,"props":1733,"children":1734},{},[1735],{"type":44,"value":1702},{"type":38,"tag":922,"props":1737,"children":1738},{},[1739],{"type":44,"value":950},{"type":38,"tag":922,"props":1741,"children":1742},{},[1743],{"type":44,"value":1711},{"type":38,"tag":870,"props":1745,"children":1746},{},[1747,1756,1760,1764,1768,1772],{"type":38,"tag":922,"props":1748,"children":1749},{},[1750],{"type":38,"tag":137,"props":1751,"children":1753},{"className":1752},[],[1754],{"type":44,"value":1755},"qwen3_600m_sft_128k_config",{"type":38,"tag":922,"props":1757,"children":1758},{},[1759],{"type":44,"value":1310},{"type":38,"tag":922,"props":1761,"children":1762},{},[1763],{"type":44,"value":945},{"type":38,"tag":922,"props":1765,"children":1766},{},[1767],{"type":44,"value":945},{"type":38,"tag":922,"props":1769,"children":1770},{},[1771],{"type":44,"value":1084},{"type":38,"tag":922,"props":1773,"children":1774},{},[1775],{"type":44,"value":1776},"600M (128K seq)",{"type":38,"tag":870,"props":1778,"children":1779},{},[1780,1789,1793,1797,1801,1805],{"type":38,"tag":922,"props":1781,"children":1782},{},[1783],{"type":38,"tag":137,"props":1784,"children":1786},{"className":1785},[],[1787],{"type":44,"value":1788},"qwen3_*_peft_config",{"type":38,"tag":922,"props":1790,"children":1791},{},[1792],{"type":44,"value":1431},{"type":38,"tag":922,"props":1794,"children":1795},{},[1796],{"type":44,"value":945},{"type":38,"tag":922,"props":1798,"children":1799},{},[1800],{"type":44,"value":945},{"type":38,"tag":922,"props":1802,"children":1803},{},[1804],{"type":44,"value":950},{"type":38,"tag":922,"props":1806,"children":1807},{},[1808],{"type":44,"value":1711},{"type":38,"tag":250,"props":1810,"children":1812},{"id":1811},"qwen3-moe",[1813],{"type":44,"value":1814},"Qwen3 MoE",{"type":38,"tag":862,"props":1816,"children":1817},{},[1818,1854],{"type":38,"tag":866,"props":1819,"children":1820},{},[1821],{"type":38,"tag":870,"props":1822,"children":1823},{},[1824,1828,1832,1836,1840,1845,1849],{"type":38,"tag":874,"props":1825,"children":1826},{},[1827],{"type":44,"value":878},{"type":38,"tag":874,"props":1829,"children":1830},{},[1831],{"type":44,"value":883},{"type":38,"tag":874,"props":1833,"children":1834},{},[1835],{"type":44,"value":888},{"type":38,"tag":874,"props":1837,"children":1838},{},[1839],{"type":44,"value":893},{"type":38,"tag":874,"props":1841,"children":1842},{},[1843],{"type":44,"value":1844},"EP",{"type":38,"tag":874,"props":1846,"children":1847},{},[1848],{"type":44,"value":898},{"type":38,"tag":874,"props":1850,"children":1851},{},[1852],{"type":44,"value":1853},"GPUs",{"type":38,"tag":915,"props":1855,"children":1856},{},[1857,1893,1929,1965,2002,2039],{"type":38,"tag":870,"props":1858,"children":1859},{},[1860,1869,1873,1877,1881,1885,1889],{"type":38,"tag":922,"props":1861,"children":1862},{},[1863],{"type":38,"tag":137,"props":1864,"children":1866},{"className":1865},[],[1867],{"type":44,"value":1868},"qwen3_30b_a3b_pretrain_config",{"type":38,"tag":922,"props":1870,"children":1871},{},[1872],{"type":44,"value":935},{"type":38,"tag":922,"props":1874,"children":1875},{},[1876],{"type":44,"value":945},{"type":38,"tag":922,"props":1878,"children":1879},{},[1880],{"type":44,"value":945},{"type":38,"tag":922,"props":1882,"children":1883},{},[1884],{"type":44,"value":1084},{"type":38,"tag":922,"props":1886,"children":1887},{},[1888],{"type":44,"value":950},{"type":38,"tag":922,"props":1890,"children":1891},{},[1892],{"type":44,"value":1084},{"type":38,"tag":870,"props":1894,"children":1895},{},[1896,1905,1909,1913,1917,1921,1925],{"type":38,"tag":922,"props":1897,"children":1898},{},[1899],{"type":38,"tag":137,"props":1900,"children":1902},{"className":1901},[],[1903],{"type":44,"value":1904},"qwen3_30b_a3b_sft_config",{"type":38,"tag":922,"props":1906,"children":1907},{},[1908],{"type":44,"value":1310},{"type":38,"tag":922,"props":1910,"children":1911},{},[1912],{"type":44,"value":945},{"type":38,"tag":922,"props":1914,"children":1915},{},[1916],{"type":44,"value":945},{"type":38,"tag":922,"props":1918,"children":1919},{},[1920],{"type":44,"value":1084},{"type":38,"tag":922,"props":1922,"children":1923},{},[1924],{"type":44,"value":950},{"type":38,"tag":922,"props":1926,"children":1927},{},[1928],{"type":44,"value":1084},{"type":38,"tag":870,"props":1930,"children":1931},{},[1932,1941,1945,1949,1953,1957,1961],{"type":38,"tag":922,"props":1933,"children":1934},{},[1935],{"type":38,"tag":137,"props":1936,"children":1938},{"className":1937},[],[1939],{"type":44,"value":1940},"qwen3_30b_a3b_peft_config",{"type":38,"tag":922,"props":1942,"children":1943},{},[1944],{"type":44,"value":1431},{"type":38,"tag":922,"props":1946,"children":1947},{},[1948],{"type":44,"value":945},{"type":38,"tag":922,"props":1950,"children":1951},{},[1952],{"type":44,"value":945},{"type":38,"tag":922,"props":1954,"children":1955},{},[1956],{"type":44,"value":945},{"type":38,"tag":922,"props":1958,"children":1959},{},[1960],{"type":44,"value":950},{"type":38,"tag":922,"props":1962,"children":1963},{},[1964],{"type":44,"value":945},{"type":38,"tag":870,"props":1966,"children":1967},{},[1968,1977,1981,1985,1989,1993,1997],{"type":38,"tag":922,"props":1969,"children":1970},{},[1971],{"type":38,"tag":137,"props":1972,"children":1974},{"className":1973},[],[1975],{"type":44,"value":1976},"qwen3_235b_a22b_pretrain_config",{"type":38,"tag":922,"props":1978,"children":1979},{},[1980],{"type":44,"value":935},{"type":38,"tag":922,"props":1982,"children":1983},{},[1984],{"type":44,"value":1042},{"type":38,"tag":922,"props":1986,"children":1987},{},[1988],{"type":44,"value":1126},{"type":38,"tag":922,"props":1990,"children":1991},{},[1992],{"type":44,"value":1084},{"type":38,"tag":922,"props":1994,"children":1995},{},[1996],{"type":44,"value":940},{"type":38,"tag":922,"props":1998,"children":1999},{},[2000],{"type":44,"value":2001},"512+",{"type":38,"tag":870,"props":2003,"children":2004},{},[2005,2014,2018,2022,2026,2030,2034],{"type":38,"tag":922,"props":2006,"children":2007},{},[2008],{"type":38,"tag":137,"props":2009,"children":2011},{"className":2010},[],[2012],{"type":44,"value":2013},"qwen3_235b_a22b_sft_config",{"type":38,"tag":922,"props":2015,"children":2016},{},[2017],{"type":44,"value":1310},{"type":38,"tag":922,"props":2019,"children":2020},{},[2021],{"type":44,"value":1042},{"type":38,"tag":922,"props":2023,"children":2024},{},[2025],{"type":44,"value":1084},{"type":38,"tag":922,"props":2027,"children":2028},{},[2029],{"type":44,"value":1084},{"type":38,"tag":922,"props":2031,"children":2032},{},[2033],{"type":44,"value":950},{"type":38,"tag":922,"props":2035,"children":2036},{},[2037],{"type":44,"value":2038},"256",{"type":38,"tag":870,"props":2040,"children":2041},{},[2042,2051,2055,2059,2063,2067,2071],{"type":38,"tag":922,"props":2043,"children":2044},{},[2045],{"type":38,"tag":137,"props":2046,"children":2048},{"className":2047},[],[2049],{"type":44,"value":2050},"qwen3_235b_a22b_peft_config",{"type":38,"tag":922,"props":2052,"children":2053},{},[2054],{"type":44,"value":1431},{"type":38,"tag":922,"props":2056,"children":2057},{},[2058],{"type":44,"value":945},{"type":38,"tag":922,"props":2060,"children":2061},{},[2062],{"type":44,"value":1042},{"type":38,"tag":922,"props":2064,"children":2065},{},[2066],{"type":44,"value":1042},{"type":38,"tag":922,"props":2068,"children":2069},{},[2070],{"type":44,"value":950},{"type":38,"tag":922,"props":2072,"children":2073},{},[2074],{"type":44,"value":1126},{"type":38,"tag":250,"props":2076,"children":2078},{"id":2077},"qwen3-next",[2079],{"type":44,"value":2080},"Qwen3-Next",{"type":38,"tag":862,"props":2082,"children":2083},{},[2084,2110],{"type":38,"tag":866,"props":2085,"children":2086},{},[2087],{"type":38,"tag":870,"props":2088,"children":2089},{},[2090,2094,2098,2102,2106],{"type":38,"tag":874,"props":2091,"children":2092},{},[2093],{"type":44,"value":878},{"type":38,"tag":874,"props":2095,"children":2096},{},[2097],{"type":44,"value":883},{"type":38,"tag":874,"props":2099,"children":2100},{},[2101],{"type":44,"value":888},{"type":38,"tag":874,"props":2103,"children":2104},{},[2105],{"type":44,"value":893},{"type":38,"tag":874,"props":2107,"children":2108},{},[2109],{"type":44,"value":1844},{"type":38,"tag":915,"props":2111,"children":2112},{},[2113,2141,2169],{"type":38,"tag":870,"props":2114,"children":2115},{},[2116,2125,2129,2133,2137],{"type":38,"tag":922,"props":2117,"children":2118},{},[2119],{"type":38,"tag":137,"props":2120,"children":2122},{"className":2121},[],[2123],{"type":44,"value":2124},"qwen3_next_80b_a3b_pretrain_config",{"type":38,"tag":922,"props":2126,"children":2127},{},[2128],{"type":44,"value":935},{"type":38,"tag":922,"props":2130,"children":2131},{},[2132],{"type":44,"value":945},{"type":38,"tag":922,"props":2134,"children":2135},{},[2136],{"type":44,"value":1042},{"type":38,"tag":922,"props":2138,"children":2139},{},[2140],{"type":44,"value":1084},{"type":38,"tag":870,"props":2142,"children":2143},{},[2144,2153,2157,2161,2165],{"type":38,"tag":922,"props":2145,"children":2146},{},[2147],{"type":38,"tag":137,"props":2148,"children":2150},{"className":2149},[],[2151],{"type":44,"value":2152},"qwen3_next_80b_a3b_sft_config",{"type":38,"tag":922,"props":2154,"children":2155},{},[2156],{"type":44,"value":1310},{"type":38,"tag":922,"props":2158,"children":2159},{},[2160],{"type":44,"value":945},{"type":38,"tag":922,"props":2162,"children":2163},{},[2164],{"type":44,"value":940},{"type":38,"tag":922,"props":2166,"children":2167},{},[2168],{"type":44,"value":1084},{"type":38,"tag":870,"props":2170,"children":2171},{},[2172,2181,2185,2189,2193],{"type":38,"tag":922,"props":2173,"children":2174},{},[2175],{"type":38,"tag":137,"props":2176,"children":2178},{"className":2177},[],[2179],{"type":44,"value":2180},"qwen3_next_80b_a3b_peft_config",{"type":38,"tag":922,"props":2182,"children":2183},{},[2184],{"type":44,"value":1431},{"type":38,"tag":922,"props":2186,"children":2187},{},[2188],{"type":44,"value":945},{"type":38,"tag":922,"props":2190,"children":2191},{},[2192],{"type":44,"value":945},{"type":38,"tag":922,"props":2194,"children":2195},{},[2196],{"type":44,"value":1042},{"type":38,"tag":250,"props":2198,"children":2200},{"id":2199},"deepseek",[2201],{"type":44,"value":2202},"DeepSeek",{"type":38,"tag":862,"props":2204,"children":2205},{},[2206,2236],{"type":38,"tag":866,"props":2207,"children":2208},{},[2209],{"type":38,"tag":870,"props":2210,"children":2211},{},[2212,2216,2220,2224,2228,2232],{"type":38,"tag":874,"props":2213,"children":2214},{},[2215],{"type":44,"value":878},{"type":38,"tag":874,"props":2217,"children":2218},{},[2219],{"type":44,"value":883},{"type":38,"tag":874,"props":2221,"children":2222},{},[2223],{"type":44,"value":888},{"type":38,"tag":874,"props":2225,"children":2226},{},[2227],{"type":44,"value":893},{"type":38,"tag":874,"props":2229,"children":2230},{},[2231],{"type":44,"value":1844},{"type":38,"tag":874,"props":2233,"children":2234},{},[2235],{"type":44,"value":1853},{"type":38,"tag":915,"props":2237,"children":2238},{},[2239,2271,2303,2336],{"type":38,"tag":870,"props":2240,"children":2241},{},[2242,2251,2255,2259,2263,2267],{"type":38,"tag":922,"props":2243,"children":2244},{},[2245],{"type":38,"tag":137,"props":2246,"children":2248},{"className":2247},[],[2249],{"type":44,"value":2250},"deepseek_v2_lite_pretrain_config",{"type":38,"tag":922,"props":2252,"children":2253},{},[2254],{"type":44,"value":935},{"type":38,"tag":922,"props":2256,"children":2257},{},[2258],{"type":44,"value":945},{"type":38,"tag":922,"props":2260,"children":2261},{},[2262],{"type":44,"value":945},{"type":38,"tag":922,"props":2264,"children":2265},{},[2266],{"type":44,"value":1084},{"type":38,"tag":922,"props":2268,"children":2269},{},[2270],{"type":44,"value":1084},{"type":38,"tag":870,"props":2272,"children":2273},{},[2274,2283,2287,2291,2295,2299],{"type":38,"tag":922,"props":2275,"children":2276},{},[2277],{"type":38,"tag":137,"props":2278,"children":2280},{"className":2279},[],[2281],{"type":44,"value":2282},"deepseek_v2_pretrain_config",{"type":38,"tag":922,"props":2284,"children":2285},{},[2286],{"type":44,"value":935},{"type":38,"tag":922,"props":2288,"children":2289},{},[2290],{"type":44,"value":945},{"type":38,"tag":922,"props":2292,"children":2293},{},[2294],{"type":44,"value":1042},{"type":38,"tag":922,"props":2296,"children":2297},{},[2298],{"type":44,"value":1168},{"type":38,"tag":922,"props":2300,"children":2301},{},[2302],{"type":44,"value":1250},{"type":38,"tag":870,"props":2304,"children":2305},{},[2306,2315,2319,2323,2327,2331],{"type":38,"tag":922,"props":2307,"children":2308},{},[2309],{"type":38,"tag":137,"props":2310,"children":2312},{"className":2311},[],[2313],{"type":44,"value":2314},"deepseek_v3_pretrain_config",{"type":38,"tag":922,"props":2316,"children":2317},{},[2318],{"type":44,"value":935},{"type":38,"tag":922,"props":2320,"children":2321},{},[2322],{"type":44,"value":940},{"type":38,"tag":922,"props":2324,"children":2325},{},[2326],{"type":44,"value":1126},{"type":38,"tag":922,"props":2328,"children":2329},{},[2330],{"type":44,"value":1209},{"type":38,"tag":922,"props":2332,"children":2333},{},[2334],{"type":44,"value":2335},"2048",{"type":38,"tag":870,"props":2337,"children":2338},{},[2339,2348,2352,2356,2360,2364],{"type":38,"tag":922,"props":2340,"children":2341},{},[2342],{"type":38,"tag":137,"props":2343,"children":2345},{"className":2344},[],[2346],{"type":44,"value":2347},"deepseek_v3_pretrain_config_32nodes",{"type":38,"tag":922,"props":2349,"children":2350},{},[2351],{"type":44,"value":935},{"type":38,"tag":922,"props":2353,"children":2354},{},[2355],{"type":44,"value":940},{"type":38,"tag":922,"props":2357,"children":2358},{},[2359],{"type":44,"value":1084},{"type":38,"tag":922,"props":2361,"children":2362},{},[2363],{"type":44,"value":1168},{"type":38,"tag":922,"props":2365,"children":2366},{},[2367],{"type":44,"value":2038},{"type":38,"tag":250,"props":2369,"children":2371},{"id":2370},"glm-45",[2372],{"type":44,"value":2373},"GLM-4.5",{"type":38,"tag":862,"props":2375,"children":2376},{},[2377,2407],{"type":38,"tag":866,"props":2378,"children":2379},{},[2380],{"type":38,"tag":870,"props":2381,"children":2382},{},[2383,2387,2391,2395,2399,2403],{"type":38,"tag":874,"props":2384,"children":2385},{},[2386],{"type":44,"value":878},{"type":38,"tag":874,"props":2388,"children":2389},{},[2390],{"type":44,"value":883},{"type":38,"tag":874,"props":2392,"children":2393},{},[2394],{"type":44,"value":888},{"type":38,"tag":874,"props":2396,"children":2397},{},[2398],{"type":44,"value":893},{"type":38,"tag":874,"props":2400,"children":2401},{},[2402],{"type":44,"value":1844},{"type":38,"tag":874,"props":2404,"children":2405},{},[2406],{"type":44,"value":1853},{"type":38,"tag":915,"props":2408,"children":2409},{},[2410,2442,2474,2506,2538,2570],{"type":38,"tag":870,"props":2411,"children":2412},{},[2413,2422,2426,2430,2434,2438],{"type":38,"tag":922,"props":2414,"children":2415},{},[2416],{"type":38,"tag":137,"props":2417,"children":2419},{"className":2418},[],[2420],{"type":44,"value":2421},"glm45_355b_pretrain_config",{"type":38,"tag":922,"props":2423,"children":2424},{},[2425],{"type":44,"value":935},{"type":38,"tag":922,"props":2427,"children":2428},{},[2429],{"type":44,"value":940},{"type":38,"tag":922,"props":2431,"children":2432},{},[2433],{"type":44,"value":1084},{"type":38,"tag":922,"props":2435,"children":2436},{},[2437],{"type":44,"value":1126},{"type":38,"tag":922,"props":2439,"children":2440},{},[2441],{"type":44,"value":2038},{"type":38,"tag":870,"props":2443,"children":2444},{},[2445,2454,2458,2462,2466,2470],{"type":38,"tag":922,"props":2446,"children":2447},{},[2448],{"type":38,"tag":137,"props":2449,"children":2451},{"className":2450},[],[2452],{"type":44,"value":2453},"glm45_air_106b_pretrain_config",{"type":38,"tag":922,"props":2455,"children":2456},{},[2457],{"type":44,"value":935},{"type":38,"tag":922,"props":2459,"children":2460},{},[2461],{"type":44,"value":945},{"type":38,"tag":922,"props":2463,"children":2464},{},[2465],{"type":44,"value":1042},{"type":38,"tag":922,"props":2467,"children":2468},{},[2469],{"type":44,"value":1084},{"type":38,"tag":922,"props":2471,"children":2472},{},[2473],{"type":44,"value":1168},{"type":38,"tag":870,"props":2475,"children":2476},{},[2477,2486,2490,2494,2498,2502],{"type":38,"tag":922,"props":2478,"children":2479},{},[2480],{"type":38,"tag":137,"props":2481,"children":2483},{"className":2482},[],[2484],{"type":44,"value":2485},"glm45_355b_sft_config",{"type":38,"tag":922,"props":2487,"children":2488},{},[2489],{"type":44,"value":1310},{"type":38,"tag":922,"props":2491,"children":2492},{},[2493],{"type":44,"value":940},{"type":38,"tag":922,"props":2495,"children":2496},{},[2497],{"type":44,"value":1084},{"type":38,"tag":922,"props":2499,"children":2500},{},[2501],{"type":44,"value":1126},{"type":38,"tag":922,"props":2503,"children":2504},{},[2505],{"type":44,"value":2038},{"type":38,"tag":870,"props":2507,"children":2508},{},[2509,2518,2522,2526,2530,2534],{"type":38,"tag":922,"props":2510,"children":2511},{},[2512],{"type":38,"tag":137,"props":2513,"children":2515},{"className":2514},[],[2516],{"type":44,"value":2517},"glm45_air_106b_sft_config",{"type":38,"tag":922,"props":2519,"children":2520},{},[2521],{"type":44,"value":1310},{"type":38,"tag":922,"props":2523,"children":2524},{},[2525],{"type":44,"value":945},{"type":38,"tag":922,"props":2527,"children":2528},{},[2529],{"type":44,"value":1042},{"type":38,"tag":922,"props":2531,"children":2532},{},[2533],{"type":44,"value":1084},{"type":38,"tag":922,"props":2535,"children":2536},{},[2537],{"type":44,"value":1168},{"type":38,"tag":870,"props":2539,"children":2540},{},[2541,2550,2554,2558,2562,2566],{"type":38,"tag":922,"props":2542,"children":2543},{},[2544],{"type":38,"tag":137,"props":2545,"children":2547},{"className":2546},[],[2548],{"type":44,"value":2549},"glm45_355b_peft_config",{"type":38,"tag":922,"props":2551,"children":2552},{},[2553],{"type":44,"value":1431},{"type":38,"tag":922,"props":2555,"children":2556},{},[2557],{"type":44,"value":940},{"type":38,"tag":922,"props":2559,"children":2560},{},[2561],{"type":44,"value":1042},{"type":38,"tag":922,"props":2563,"children":2564},{},[2565],{"type":44,"value":1042},{"type":38,"tag":922,"props":2567,"children":2568},{},[2569],{"type":44,"value":1168},{"type":38,"tag":870,"props":2571,"children":2572},{},[2573,2582,2586,2590,2594,2598],{"type":38,"tag":922,"props":2574,"children":2575},{},[2576],{"type":38,"tag":137,"props":2577,"children":2579},{"className":2578},[],[2580],{"type":44,"value":2581},"glm45_air_106b_peft_config",{"type":38,"tag":922,"props":2583,"children":2584},{},[2585],{"type":44,"value":1431},{"type":38,"tag":922,"props":2587,"children":2588},{},[2589],{"type":44,"value":945},{"type":38,"tag":922,"props":2591,"children":2592},{},[2593],{"type":44,"value":940},{"type":38,"tag":922,"props":2595,"children":2596},{},[2597],{"type":44,"value":1042},{"type":38,"tag":922,"props":2599,"children":2600},{},[2601],{"type":44,"value":1084},{"type":38,"tag":250,"props":2603,"children":2605},{"id":2604},"gemma",[2606],{"type":44,"value":2607},"Gemma",{"type":38,"tag":862,"props":2609,"children":2610},{},[2611,2637],{"type":38,"tag":866,"props":2612,"children":2613},{},[2614],{"type":38,"tag":870,"props":2615,"children":2616},{},[2617,2621,2625,2629,2633],{"type":38,"tag":874,"props":2618,"children":2619},{},[2620],{"type":44,"value":878},{"type":38,"tag":874,"props":2622,"children":2623},{},[2624],{"type":44,"value":883},{"type":38,"tag":874,"props":2626,"children":2627},{},[2628],{"type":44,"value":888},{"type":38,"tag":874,"props":2630,"children":2631},{},[2632],{"type":44,"value":893},{"type":38,"tag":874,"props":2634,"children":2635},{},[2636],{"type":44,"value":1571},{"type":38,"tag":915,"props":2638,"children":2639},{},[2640,2670],{"type":38,"tag":870,"props":2641,"children":2642},{},[2643,2652,2656,2661,2665],{"type":38,"tag":922,"props":2644,"children":2645},{},[2646],{"type":38,"tag":137,"props":2647,"children":2649},{"className":2648},[],[2650],{"type":44,"value":2651},"gemma2_*_{pretrain,sft,peft}_config",{"type":38,"tag":922,"props":2653,"children":2654},{},[2655],{"type":44,"value":1591},{"type":38,"tag":922,"props":2657,"children":2658},{},[2659],{"type":44,"value":2660},"2–8",{"type":38,"tag":922,"props":2662,"children":2663},{},[2664],{"type":44,"value":1702},{"type":38,"tag":922,"props":2666,"children":2667},{},[2668],{"type":44,"value":2669},"2B, 9B, 27B",{"type":38,"tag":870,"props":2671,"children":2672},{},[2673,2682,2686,2690,2694],{"type":38,"tag":922,"props":2674,"children":2675},{},[2676],{"type":38,"tag":137,"props":2677,"children":2679},{"className":2678},[],[2680],{"type":44,"value":2681},"gemma3_1b_{pretrain,sft,peft}_config",{"type":38,"tag":922,"props":2683,"children":2684},{},[2685],{"type":44,"value":1591},{"type":38,"tag":922,"props":2687,"children":2688},{},[2689],{"type":44,"value":945},{"type":38,"tag":922,"props":2691,"children":2692},{},[2693],{"type":44,"value":945},{"type":38,"tag":922,"props":2695,"children":2696},{},[2697],{"type":44,"value":2698},"1B (32K seq)",{"type":38,"tag":250,"props":2700,"children":2702},{"id":2701},"nemotronh-nemotron",[2703],{"type":44,"value":2704},"NemotronH \u002F Nemotron",{"type":38,"tag":862,"props":2706,"children":2707},{},[2708,2739],{"type":38,"tag":866,"props":2709,"children":2710},{},[2711],{"type":38,"tag":870,"props":2712,"children":2713},{},[2714,2718,2722,2726,2730,2734],{"type":38,"tag":874,"props":2715,"children":2716},{},[2717],{"type":44,"value":878},{"type":38,"tag":874,"props":2719,"children":2720},{},[2721],{"type":44,"value":883},{"type":38,"tag":874,"props":2723,"children":2724},{},[2725],{"type":44,"value":888},{"type":38,"tag":874,"props":2727,"children":2728},{},[2729],{"type":44,"value":893},{"type":38,"tag":874,"props":2731,"children":2732},{},[2733],{"type":44,"value":1844},{"type":38,"tag":874,"props":2735,"children":2736},{},[2737],{"type":44,"value":2738},"Notes",{"type":38,"tag":915,"props":2740,"children":2741},{},[2742,2776,2810,2843],{"type":38,"tag":870,"props":2743,"children":2744},{},[2745,2754,2759,2763,2767,2771],{"type":38,"tag":922,"props":2746,"children":2747},{},[2748],{"type":38,"tag":137,"props":2749,"children":2751},{"className":2750},[],[2752],{"type":44,"value":2753},"nemotronh_{4b,8b,47b,56b}_*_config",{"type":38,"tag":922,"props":2755,"children":2756},{},[2757],{"type":44,"value":2758},"P\u002FS\u002FPEFT",{"type":38,"tag":922,"props":2760,"children":2761},{},[2762],{"type":44,"value":1596},{"type":38,"tag":922,"props":2764,"children":2765},{},[2766],{"type":44,"value":1601},{"type":38,"tag":922,"props":2768,"children":2769},{},[2770],{"type":44,"value":950},{"type":38,"tag":922,"props":2772,"children":2773},{},[2774],{"type":44,"value":2775},"Dense SSM-hybrid",{"type":38,"tag":870,"props":2777,"children":2778},{},[2779,2788,2792,2797,2801,2805],{"type":38,"tag":922,"props":2780,"children":2781},{},[2782],{"type":38,"tag":137,"props":2783,"children":2785},{"className":2784},[],[2786],{"type":44,"value":2787},"nemotron_3_nano_*_config",{"type":38,"tag":922,"props":2789,"children":2790},{},[2791],{"type":44,"value":2758},{"type":38,"tag":922,"props":2793,"children":2794},{},[2795],{"type":44,"value":2796},"varies",{"type":38,"tag":922,"props":2798,"children":2799},{},[2800],{"type":44,"value":945},{"type":38,"tag":922,"props":2802,"children":2803},{},[2804],{"type":44,"value":1084},{"type":38,"tag":922,"props":2806,"children":2807},{},[2808],{"type":44,"value":2809},"MoE + Mamba",{"type":38,"tag":870,"props":2811,"children":2812},{},[2813,2822,2826,2830,2834,2838],{"type":38,"tag":922,"props":2814,"children":2815},{},[2816],{"type":38,"tag":137,"props":2817,"children":2819},{"className":2818},[],[2820],{"type":44,"value":2821},"nemotron_3_super_*_config",{"type":38,"tag":922,"props":2823,"children":2824},{},[2825],{"type":44,"value":2758},{"type":38,"tag":922,"props":2827,"children":2828},{},[2829],{"type":44,"value":1042},{"type":38,"tag":922,"props":2831,"children":2832},{},[2833],{"type":44,"value":945},{"type":38,"tag":922,"props":2835,"children":2836},{},[2837],{"type":44,"value":1084},{"type":38,"tag":922,"props":2839,"children":2840},{},[2841],{"type":44,"value":2842},"MoE + Mamba, ~40% CUDA graph gain",{"type":38,"tag":870,"props":2844,"children":2845},{},[2846,2855,2859,2863,2867,2871],{"type":38,"tag":922,"props":2847,"children":2848},{},[2849],{"type":38,"tag":137,"props":2850,"children":2852},{"className":2851},[],[2853],{"type":44,"value":2854},"nemotron_nano_{9b,12b}_v2_*_config",{"type":38,"tag":922,"props":2856,"children":2857},{},[2858],{"type":44,"value":2758},{"type":38,"tag":922,"props":2860,"children":2861},{},[2862],{"type":44,"value":2796},{"type":38,"tag":922,"props":2864,"children":2865},{},[2866],{"type":44,"value":945},{"type":38,"tag":922,"props":2868,"children":2869},{},[2870],{"type":44,"value":950},{"type":38,"tag":922,"props":2872,"children":2873},{},[2874],{"type":44,"value":2875},"Dense",{"type":38,"tag":250,"props":2877,"children":2879},{"id":2878},"other-models",[2880],{"type":44,"value":2881},"Other Models",{"type":38,"tag":862,"props":2883,"children":2884},{},[2885,2903],{"type":38,"tag":866,"props":2886,"children":2887},{},[2888],{"type":38,"tag":870,"props":2889,"children":2890},{},[2891,2895,2899],{"type":38,"tag":874,"props":2892,"children":2893},{},[2894],{"type":44,"value":878},{"type":38,"tag":874,"props":2896,"children":2897},{},[2898],{"type":44,"value":883},{"type":38,"tag":874,"props":2900,"children":2901},{},[2902],{"type":44,"value":2738},{"type":38,"tag":915,"props":2904,"children":2905},{},[2906,2927,2947,2968,2989,3010,3031,3052],{"type":38,"tag":870,"props":2907,"children":2908},{},[2909,2918,2922],{"type":38,"tag":922,"props":2910,"children":2911},{},[2912],{"type":38,"tag":137,"props":2913,"children":2915},{"className":2914},[],[2916],{"type":44,"value":2917},"moonlight_16b_{pretrain,sft,peft}_config",{"type":38,"tag":922,"props":2919,"children":2920},{},[2921],{"type":44,"value":1591},{"type":38,"tag":922,"props":2923,"children":2924},{},[2925],{"type":44,"value":2926},"MoE EP=8",{"type":38,"tag":870,"props":2928,"children":2929},{},[2930,2939,2943],{"type":38,"tag":922,"props":2931,"children":2932},{},[2933],{"type":38,"tag":137,"props":2934,"children":2936},{"className":2935},[],[2937],{"type":44,"value":2938},"olmoe_7b_{pretrain,sft,peft}_config",{"type":38,"tag":922,"props":2940,"children":2941},{},[2942],{"type":44,"value":1591},{"type":38,"tag":922,"props":2944,"children":2945},{},[2946],{"type":44,"value":2926},{"type":38,"tag":870,"props":2948,"children":2949},{},[2950,2959,2964],{"type":38,"tag":922,"props":2951,"children":2952},{},[2953],{"type":38,"tag":137,"props":2954,"children":2956},{"className":2955},[],[2957],{"type":44,"value":2958},"ministral3_{3b,8b,14b}_{sft,peft}_config",{"type":38,"tag":922,"props":2960,"children":2961},{},[2962],{"type":44,"value":2963},"SFT\u002FPEFT",{"type":38,"tag":922,"props":2965,"children":2966},{},[2967],{"type":44,"value":2875},{"type":38,"tag":870,"props":2969,"children":2970},{},[2971,2980,2984],{"type":38,"tag":922,"props":2972,"children":2973},{},[2974],{"type":38,"tag":137,"props":2975,"children":2977},{"className":2976},[],[2978],{"type":44,"value":2979},"gpt_oss_20b_*_config",{"type":38,"tag":922,"props":2981,"children":2982},{},[2983],{"type":44,"value":1591},{"type":38,"tag":922,"props":2985,"children":2986},{},[2987],{"type":44,"value":2988},"MoE + FP8\u002FMXFP8 variants",{"type":38,"tag":870,"props":2990,"children":2991},{},[2992,3001,3005],{"type":38,"tag":922,"props":2993,"children":2994},{},[2995],{"type":38,"tag":137,"props":2996,"children":2998},{"className":2997},[],[2999],{"type":44,"value":3000},"gpt_oss_120b_*_config",{"type":38,"tag":922,"props":3002,"children":3003},{},[3004],{"type":44,"value":1591},{"type":38,"tag":922,"props":3006,"children":3007},{},[3008],{"type":44,"value":3009},"MoE",{"type":38,"tag":870,"props":3011,"children":3012},{},[3013,3022,3026],{"type":38,"tag":922,"props":3014,"children":3015},{},[3016],{"type":38,"tag":137,"props":3017,"children":3019},{"className":3018},[],[3020],{"type":44,"value":3021},"vanilla_gpt_pretrain_config",{"type":38,"tag":922,"props":3023,"children":3024},{},[3025],{"type":44,"value":935},{"type":38,"tag":922,"props":3027,"children":3028},{},[3029],{"type":44,"value":3030},"MLM\u002FBridge parity baseline",{"type":38,"tag":870,"props":3032,"children":3033},{},[3034,3043,3047],{"type":38,"tag":922,"props":3035,"children":3036},{},[3037],{"type":38,"tag":137,"props":3038,"children":3040},{"className":3039},[],[3041],{"type":44,"value":3042},"gpt3_175b_pretrain_config",{"type":38,"tag":922,"props":3044,"children":3045},{},[3046],{"type":44,"value":935},{"type":38,"tag":922,"props":3048,"children":3049},{},[3050],{"type":44,"value":3051},"TP=4, PP=8, VP=6",{"type":38,"tag":870,"props":3053,"children":3054},{},[3055,3064,3068],{"type":38,"tag":922,"props":3056,"children":3057},{},[3058],{"type":38,"tag":137,"props":3059,"children":3061},{"className":3060},[],[3062],{"type":44,"value":3063},"kimi_k2_pretrain_config",{"type":38,"tag":922,"props":3065,"children":3066},{},[3067],{"type":44,"value":935},{"type":38,"tag":922,"props":3069,"children":3070},{},[3071],{"type":44,"value":3072},"1T MoE, TP=2 PP=16 EP=32",{"type":38,"tag":250,"props":3074,"children":3076},{"id":3075},"vlm-recipes",[3077],{"type":44,"value":3078},"VLM Recipes",{"type":38,"tag":862,"props":3080,"children":3081},{},[3082,3112],{"type":38,"tag":866,"props":3083,"children":3084},{},[3085],{"type":38,"tag":870,"props":3086,"children":3087},{},[3088,3092,3096,3100,3104,3108],{"type":38,"tag":874,"props":3089,"children":3090},{},[3091],{"type":44,"value":878},{"type":38,"tag":874,"props":3093,"children":3094},{},[3095],{"type":44,"value":883},{"type":38,"tag":874,"props":3097,"children":3098},{},[3099],{"type":44,"value":888},{"type":38,"tag":874,"props":3101,"children":3102},{},[3103],{"type":44,"value":893},{"type":38,"tag":874,"props":3105,"children":3106},{},[3107],{"type":44,"value":1844},{"type":38,"tag":874,"props":3109,"children":3110},{},[3111],{"type":44,"value":1853},{"type":38,"tag":915,"props":3113,"children":3114},{},[3115,3148,3181,3214,3246,3280],{"type":38,"tag":870,"props":3116,"children":3117},{},[3118,3127,3131,3135,3139,3143],{"type":38,"tag":922,"props":3119,"children":3120},{},[3121],{"type":38,"tag":137,"props":3122,"children":3124},{"className":3123},[],[3125],{"type":44,"value":3126},"gemma3_vl_{4b,12b,27b}_{sft,peft}_config",{"type":38,"tag":922,"props":3128,"children":3129},{},[3130],{"type":44,"value":2963},{"type":38,"tag":922,"props":3132,"children":3133},{},[3134],{"type":44,"value":1596},{"type":38,"tag":922,"props":3136,"children":3137},{},[3138],{"type":44,"value":1702},{"type":38,"tag":922,"props":3140,"children":3141},{},[3142],{"type":44,"value":950},{"type":38,"tag":922,"props":3144,"children":3145},{},[3146],{"type":44,"value":3147},"1–16",{"type":38,"tag":870,"props":3149,"children":3150},{},[3151,3160,3164,3168,3172,3176],{"type":38,"tag":922,"props":3152,"children":3153},{},[3154],{"type":38,"tag":137,"props":3155,"children":3157},{"className":3156},[],[3158],{"type":44,"value":3159},"qwen25_vl_{3b,7b,32b,72b}_{sft,peft}_config",{"type":38,"tag":922,"props":3161,"children":3162},{},[3163],{"type":44,"value":2963},{"type":38,"tag":922,"props":3165,"children":3166},{},[3167],{"type":44,"value":1596},{"type":38,"tag":922,"props":3169,"children":3170},{},[3171],{"type":44,"value":1601},{"type":38,"tag":922,"props":3173,"children":3174},{},[3175],{"type":44,"value":950},{"type":38,"tag":922,"props":3177,"children":3178},{},[3179],{"type":44,"value":3180},"1–32",{"type":38,"tag":870,"props":3182,"children":3183},{},[3184,3193,3197,3201,3205,3209],{"type":38,"tag":922,"props":3185,"children":3186},{},[3187],{"type":38,"tag":137,"props":3188,"children":3190},{"className":3189},[],[3191],{"type":44,"value":3192},"qwen3_vl_{8b,30b_a3b,235b_a22b}_{sft,peft}_config",{"type":38,"tag":922,"props":3194,"children":3195},{},[3196],{"type":44,"value":2963},{"type":38,"tag":922,"props":3198,"children":3199},{},[3200],{"type":44,"value":1601},{"type":38,"tag":922,"props":3202,"children":3203},{},[3204],{"type":44,"value":1596},{"type":38,"tag":922,"props":3206,"children":3207},{},[3208],{"type":44,"value":3180},{"type":38,"tag":922,"props":3210,"children":3211},{},[3212],{"type":44,"value":3213},"1–512",{"type":38,"tag":870,"props":3215,"children":3216},{},[3217,3226,3230,3234,3238,3242],{"type":38,"tag":922,"props":3218,"children":3219},{},[3220],{"type":38,"tag":137,"props":3221,"children":3223},{"className":3222},[],[3224],{"type":44,"value":3225},"qwen35_vl_*_{sft,peft}_config",{"type":38,"tag":922,"props":3227,"children":3228},{},[3229],{"type":44,"value":2963},{"type":38,"tag":922,"props":3231,"children":3232},{},[3233],{"type":44,"value":2796},{"type":38,"tag":922,"props":3235,"children":3236},{},[3237],{"type":44,"value":2796},{"type":38,"tag":922,"props":3239,"children":3240},{},[3241],{"type":44,"value":2796},{"type":38,"tag":922,"props":3243,"children":3244},{},[3245],{"type":44,"value":2796},{"type":38,"tag":870,"props":3247,"children":3248},{},[3249,3258,3262,3266,3270,3275],{"type":38,"tag":922,"props":3250,"children":3251},{},[3252],{"type":38,"tag":137,"props":3253,"children":3255},{"className":3254},[],[3256],{"type":44,"value":3257},"glm_45v_{sft,peft}_config",{"type":38,"tag":922,"props":3259,"children":3260},{},[3261],{"type":44,"value":2963},{"type":38,"tag":922,"props":3263,"children":3264},{},[3265],{"type":44,"value":945},{"type":38,"tag":922,"props":3267,"children":3268},{},[3269],{"type":44,"value":1084},{"type":38,"tag":922,"props":3271,"children":3272},{},[3273],{"type":44,"value":3274},"4–16",{"type":38,"tag":922,"props":3276,"children":3277},{},[3278],{"type":44,"value":3279},"64–512",{"type":38,"tag":870,"props":3281,"children":3282},{},[3283,3292,3296,3301,3305,3309],{"type":38,"tag":922,"props":3284,"children":3285},{},[3286],{"type":38,"tag":137,"props":3287,"children":3289},{"className":3288},[],[3290],{"type":44,"value":3291},"nemotron_nano_v2_vl_12b_{sft,peft}_config",{"type":38,"tag":922,"props":3293,"children":3294},{},[3295],{"type":44,"value":2963},{"type":38,"tag":922,"props":3297,"children":3298},{},[3299],{"type":44,"value":3300},"2–4",{"type":38,"tag":922,"props":3302,"children":3303},{},[3304],{"type":44,"value":945},{"type":38,"tag":922,"props":3306,"children":3307},{},[3308],{"type":44,"value":950},{"type":38,"tag":922,"props":3310,"children":3311},{},[3312],{"type":44,"value":1084},{"type":38,"tag":250,"props":3314,"children":3316},{"id":3315},"diffusion-recipes",[3317],{"type":44,"value":3318},"Diffusion Recipes",{"type":38,"tag":862,"props":3320,"children":3321},{},[3322,3344],{"type":38,"tag":866,"props":3323,"children":3324},{},[3325],{"type":38,"tag":870,"props":3326,"children":3327},{},[3328,3332,3336,3340],{"type":38,"tag":874,"props":3329,"children":3330},{},[3331],{"type":44,"value":878},{"type":38,"tag":874,"props":3333,"children":3334},{},[3335],{"type":44,"value":883},{"type":38,"tag":874,"props":3337,"children":3338},{},[3339],{"type":44,"value":888},{"type":38,"tag":874,"props":3341,"children":3342},{},[3343],{"type":44,"value":898},{"type":38,"tag":915,"props":3345,"children":3346},{},[3347,3372,3396],{"type":38,"tag":870,"props":3348,"children":3349},{},[3350,3359,3364,3368],{"type":38,"tag":922,"props":3351,"children":3352},{},[3353],{"type":38,"tag":137,"props":3354,"children":3356},{"className":3355},[],[3357],{"type":44,"value":3358},"wan_1_3B_{pretrain,sft}_config",{"type":38,"tag":922,"props":3360,"children":3361},{},[3362],{"type":44,"value":3363},"P\u002FSFT",{"type":38,"tag":922,"props":3365,"children":3366},{},[3367],{"type":44,"value":945},{"type":38,"tag":922,"props":3369,"children":3370},{},[3371],{"type":44,"value":1084},{"type":38,"tag":870,"props":3373,"children":3374},{},[3375,3384,3388,3392],{"type":38,"tag":922,"props":3376,"children":3377},{},[3378],{"type":38,"tag":137,"props":3379,"children":3381},{"className":3380},[],[3382],{"type":44,"value":3383},"wan_14B_{pretrain,sft}_config",{"type":38,"tag":922,"props":3385,"children":3386},{},[3387],{"type":44,"value":3363},{"type":38,"tag":922,"props":3389,"children":3390},{},[3391],{"type":44,"value":940},{"type":38,"tag":922,"props":3393,"children":3394},{},[3395],{"type":44,"value":1042},{"type":38,"tag":870,"props":3397,"children":3398},{},[3399,3408,3412,3416],{"type":38,"tag":922,"props":3400,"children":3401},{},[3402],{"type":38,"tag":137,"props":3403,"children":3405},{"className":3404},[],[3406],{"type":44,"value":3407},"flux_12b_{pretrain,sft}_config",{"type":38,"tag":922,"props":3409,"children":3410},{},[3411],{"type":44,"value":3363},{"type":38,"tag":922,"props":3413,"children":3414},{},[3415],{"type":44,"value":940},{"type":38,"tag":922,"props":3417,"children":3418},{},[3419],{"type":44,"value":945},{"type":38,"tag":240,"props":3421,"children":3422},{},[],{"type":38,"tag":53,"props":3424,"children":3426},{"id":3425},"performance-recipe-index",[3427],{"type":44,"value":3428},"Performance Recipe Index",{"type":38,"tag":47,"props":3430,"children":3431},{},[3432,3434,3440,3442,3447],{"type":44,"value":3433},"Perf recipe source lives under ",{"type":38,"tag":137,"props":3435,"children":3437},{"className":3436},[],[3438],{"type":44,"value":3439},"src\u002Fmegatron\u002Fbridge\u002Fperf_recipes\u002F",{"type":44,"value":3441},". The\nperformance launcher in ",{"type":38,"tag":137,"props":3443,"children":3445},{"className":3444},[],[3446],{"type":44,"value":167},{"type":44,"value":3448}," resolves those flat recipe names\nand derives compatibility workload views from the selected flat recipe when\nlegacy helper paths still need them.",{"type":38,"tag":3450,"props":3451,"children":3452},"blockquote",{},[3453],{"type":38,"tag":47,"props":3454,"children":3455},{},[3456,3461,3463,3468,3470,3475,3477,3482],{"type":38,"tag":70,"props":3457,"children":3458},{},[3459],{"type":44,"value":3460},"Important:",{"type":44,"value":3462}," Perf recipes are designed for ",{"type":38,"tag":70,"props":3464,"children":3465},{},[3466],{"type":44,"value":3467},"upper-bound throughput\nbenchmarks",{"type":44,"value":3469},", not production training. They run ",{"type":38,"tag":70,"props":3471,"children":3472},{},[3473],{"type":44,"value":3474},"50 iterations",{"type":44,"value":3476}," on ",{"type":38,"tag":70,"props":3478,"children":3479},{},[3480],{"type":44,"value":3481},"mock\ndata",{"type":44,"value":3483}," by default. Throughput numbers are aspirational targets, not validated\nconvergence configs.",{"type":38,"tag":250,"props":3485,"children":3487},{"id":3486},"llama-3-31",[3488],{"type":44,"value":3489},"Llama 3 \u002F 3.1",{"type":38,"tag":862,"props":3491,"children":3492},{},[3493,3518],{"type":38,"tag":866,"props":3494,"children":3495},{},[3496],{"type":38,"tag":870,"props":3497,"children":3498},{},[3499,3504,3508,3513],{"type":38,"tag":874,"props":3500,"children":3501},{},[3502],{"type":44,"value":3503},"Model",{"type":38,"tag":874,"props":3505,"children":3506},{},[3507],{"type":44,"value":1853},{"type":38,"tag":874,"props":3509,"children":3510},{},[3511],{"type":44,"value":3512},"GPU Types",{"type":38,"tag":874,"props":3514,"children":3515},{},[3516],{"type":44,"value":3517},"Key Features",{"type":38,"tag":915,"props":3519,"children":3520},{},[3521,3543,3565],{"type":38,"tag":870,"props":3522,"children":3523},{},[3524,3529,3533,3538],{"type":38,"tag":922,"props":3525,"children":3526},{},[3527],{"type":44,"value":3528},"Llama 3 8B",{"type":38,"tag":922,"props":3530,"children":3531},{},[3532],{"type":44,"value":1084},{"type":38,"tag":922,"props":3534,"children":3535},{},[3536],{"type":44,"value":3537},"H100, B200, B300, GB200, GB300, R100",{"type":38,"tag":922,"props":3539,"children":3540},{},[3541],{"type":44,"value":3542},"CUDA graphs (local), FSDP on GB variants",{"type":38,"tag":870,"props":3544,"children":3545},{},[3546,3551,3555,3560],{"type":38,"tag":922,"props":3547,"children":3548},{},[3549],{"type":44,"value":3550},"Llama 3 70B",{"type":38,"tag":922,"props":3552,"children":3553},{},[3554],{"type":44,"value":1209},{"type":38,"tag":922,"props":3556,"children":3557},{},[3558],{"type":44,"value":3559},"H100, B200, B300, GB200, GB300",{"type":38,"tag":922,"props":3561,"children":3562},{},[3563],{"type":44,"value":3564},"TP comm overlap (userbuffers), FSDP, CUDA graphs",{"type":38,"tag":870,"props":3566,"children":3567},{},[3568,3573,3578,3582],{"type":38,"tag":922,"props":3569,"children":3570},{},[3571],{"type":44,"value":3572},"Llama 3.1 405B",{"type":38,"tag":922,"props":3574,"children":3575},{},[3576],{"type":44,"value":3577},"128–1024",{"type":38,"tag":922,"props":3579,"children":3580},{},[3581],{"type":44,"value":3559},{"type":38,"tag":922,"props":3583,"children":3584},{},[3585],{"type":44,"value":3586},"TP+CP comm overlap (userbuffers), FSDP, heavy PP\u002FVP",{"type":38,"tag":47,"props":3588,"children":3589},{},[3590],{"type":44,"value":3591},"SFT\u002FLoRA variants also exist (e.g. 8B SFT with packed sequences, 70B SFT on 32 GPUs).",{"type":38,"tag":250,"props":3593,"children":3595},{"id":3594},"deepseek-v3",[3596],{"type":44,"value":3597},"DeepSeek V3",{"type":38,"tag":862,"props":3599,"children":3600},{},[3601,3623],{"type":38,"tag":866,"props":3602,"children":3603},{},[3604],{"type":38,"tag":870,"props":3605,"children":3606},{},[3607,3611,3615,3619],{"type":38,"tag":874,"props":3608,"children":3609},{},[3610],{"type":44,"value":3503},{"type":38,"tag":874,"props":3612,"children":3613},{},[3614],{"type":44,"value":1853},{"type":38,"tag":874,"props":3616,"children":3617},{},[3618],{"type":44,"value":3512},{"type":38,"tag":874,"props":3620,"children":3621},{},[3622],{"type":44,"value":3517},{"type":38,"tag":915,"props":3624,"children":3625},{},[3626],{"type":38,"tag":870,"props":3627,"children":3628},{},[3629,3634,3639,3643],{"type":38,"tag":922,"props":3630,"children":3631},{},[3632],{"type":44,"value":3633},"DeepSeek V3 (671B MoE)",{"type":38,"tag":922,"props":3635,"children":3636},{},[3637],{"type":44,"value":3638},"256–1024",{"type":38,"tag":922,"props":3640,"children":3641},{},[3642],{"type":44,"value":3559},{"type":38,"tag":922,"props":3644,"children":3645},{},[3646],{"type":44,"value":3647},"HybridEP dispatcher, MLA recompute, CUDA graphs (TE scoped)",{"type":38,"tag":250,"props":3649,"children":3651},{"id":3650},"qwen3-moe-1",[3652],{"type":44,"value":1814},{"type":38,"tag":862,"props":3654,"children":3655},{},[3656,3678],{"type":38,"tag":866,"props":3657,"children":3658},{},[3659],{"type":38,"tag":870,"props":3660,"children":3661},{},[3662,3666,3670,3674],{"type":38,"tag":874,"props":3663,"children":3664},{},[3665],{"type":44,"value":3503},{"type":38,"tag":874,"props":3667,"children":3668},{},[3669],{"type":44,"value":1853},{"type":38,"tag":874,"props":3671,"children":3672},{},[3673],{"type":44,"value":3512},{"type":38,"tag":874,"props":3675,"children":3676},{},[3677],{"type":44,"value":3517},{"type":38,"tag":915,"props":3679,"children":3680},{},[3681,3703,3725],{"type":38,"tag":870,"props":3682,"children":3683},{},[3684,3689,3694,3698],{"type":38,"tag":922,"props":3685,"children":3686},{},[3687],{"type":44,"value":3688},"Qwen3 30B-A3B",{"type":38,"tag":922,"props":3690,"children":3691},{},[3692],{"type":44,"value":3693},"8–16",{"type":38,"tag":922,"props":3695,"children":3696},{},[3697],{"type":44,"value":3559},{"type":38,"tag":922,"props":3699,"children":3700},{},[3701],{"type":44,"value":3702},"MoE alltoall\u002Fflex dispatcher",{"type":38,"tag":870,"props":3704,"children":3705},{},[3706,3711,3716,3720],{"type":38,"tag":922,"props":3707,"children":3708},{},[3709],{"type":44,"value":3710},"Qwen3 235B-A22B",{"type":38,"tag":922,"props":3712,"children":3713},{},[3714],{"type":44,"value":3715},"64–256",{"type":38,"tag":922,"props":3717,"children":3718},{},[3719],{"type":44,"value":3559},{"type":38,"tag":922,"props":3721,"children":3722},{},[3723],{"type":44,"value":3724},"TP comm overlap, CUDA graphs, MoE a2a overlap",{"type":38,"tag":870,"props":3726,"children":3727},{},[3728,3733,3738,3742],{"type":38,"tag":922,"props":3729,"children":3730},{},[3731],{"type":44,"value":3732},"Qwen3-Next 80B-A3B",{"type":38,"tag":922,"props":3734,"children":3735},{},[3736],{"type":44,"value":3737},"64–128",{"type":38,"tag":922,"props":3739,"children":3740},{},[3741],{"type":44,"value":3559},{"type":38,"tag":922,"props":3743,"children":3744},{},[3745],{"type":44,"value":3746},"EP 64–128",{"type":38,"tag":250,"props":3748,"children":3750},{"id":3749},"qwen3-vl",[3751],{"type":44,"value":3752},"Qwen3-VL",{"type":38,"tag":862,"props":3754,"children":3755},{},[3756,3778],{"type":38,"tag":866,"props":3757,"children":3758},{},[3759],{"type":38,"tag":870,"props":3760,"children":3761},{},[3762,3766,3770,3774],{"type":38,"tag":874,"props":3763,"children":3764},{},[3765],{"type":44,"value":3503},{"type":38,"tag":874,"props":3767,"children":3768},{},[3769],{"type":44,"value":1853},{"type":38,"tag":874,"props":3771,"children":3772},{},[3773],{"type":44,"value":3512},{"type":38,"tag":874,"props":3775,"children":3776},{},[3777],{"type":44,"value":3517},{"type":38,"tag":915,"props":3779,"children":3780},{},[3781,3802],{"type":38,"tag":870,"props":3782,"children":3783},{},[3784,3789,3793,3797],{"type":38,"tag":922,"props":3785,"children":3786},{},[3787],{"type":44,"value":3788},"Qwen3-VL 30B-A3B",{"type":38,"tag":922,"props":3790,"children":3791},{},[3792],{"type":44,"value":3693},{"type":38,"tag":922,"props":3794,"children":3795},{},[3796],{"type":44,"value":3559},{"type":38,"tag":922,"props":3798,"children":3799},{},[3800],{"type":44,"value":3801},"VLM + MoE",{"type":38,"tag":870,"props":3803,"children":3804},{},[3805,3810,3814,3818],{"type":38,"tag":922,"props":3806,"children":3807},{},[3808],{"type":44,"value":3809},"Qwen3-VL 235B-A22B",{"type":38,"tag":922,"props":3811,"children":3812},{},[3813],{"type":44,"value":3715},{"type":38,"tag":922,"props":3815,"children":3816},{},[3817],{"type":44,"value":3559},{"type":38,"tag":922,"props":3819,"children":3820},{},[3821],{"type":44,"value":3822},"VLM + MoE, TP comm overlap",{"type":38,"tag":250,"props":3824,"children":3826},{"id":3825},"kimi-k2",[3827],{"type":44,"value":3828},"Kimi K2",{"type":38,"tag":862,"props":3830,"children":3831},{},[3832,3854],{"type":38,"tag":866,"props":3833,"children":3834},{},[3835],{"type":38,"tag":870,"props":3836,"children":3837},{},[3838,3842,3846,3850],{"type":38,"tag":874,"props":3839,"children":3840},{},[3841],{"type":44,"value":3503},{"type":38,"tag":874,"props":3843,"children":3844},{},[3845],{"type":44,"value":1853},{"type":38,"tag":874,"props":3847,"children":3848},{},[3849],{"type":44,"value":3512},{"type":38,"tag":874,"props":3851,"children":3852},{},[3853],{"type":44,"value":3517},{"type":38,"tag":915,"props":3855,"children":3856},{},[3857],{"type":38,"tag":870,"props":3858,"children":3859},{},[3860,3865,3869,3873],{"type":38,"tag":922,"props":3861,"children":3862},{},[3863],{"type":44,"value":3864},"Kimi K2 (1T MoE)",{"type":38,"tag":922,"props":3866,"children":3867},{},[3868],{"type":44,"value":3638},{"type":38,"tag":922,"props":3870,"children":3871},{},[3872],{"type":44,"value":3559},{"type":38,"tag":922,"props":3874,"children":3875},{},[3876],{"type":44,"value":3877},"Muon\u002FAdam optimizer, HybridEP, pipeline layout helpers",{"type":38,"tag":250,"props":3879,"children":3881},{"id":3880},"nemotronh",[3882],{"type":44,"value":3883},"NemotronH",{"type":38,"tag":862,"props":3885,"children":3886},{},[3887,3909],{"type":38,"tag":866,"props":3888,"children":3889},{},[3890],{"type":38,"tag":870,"props":3891,"children":3892},{},[3893,3897,3901,3905],{"type":38,"tag":874,"props":3894,"children":3895},{},[3896],{"type":44,"value":3503},{"type":38,"tag":874,"props":3898,"children":3899},{},[3900],{"type":44,"value":1853},{"type":38,"tag":874,"props":3902,"children":3903},{},[3904],{"type":44,"value":3512},{"type":38,"tag":874,"props":3906,"children":3907},{},[3908],{"type":44,"value":3517},{"type":38,"tag":915,"props":3910,"children":3911},{},[3912,3933,3954],{"type":38,"tag":870,"props":3913,"children":3914},{},[3915,3920,3924,3928],{"type":38,"tag":922,"props":3916,"children":3917},{},[3918],{"type":44,"value":3919},"Nemotron 3 Nano (30B MoE+Mamba)",{"type":38,"tag":922,"props":3921,"children":3922},{},[3923],{"type":44,"value":3693},{"type":38,"tag":922,"props":3925,"children":3926},{},[3927],{"type":44,"value":3559},{"type":38,"tag":922,"props":3929,"children":3930},{},[3931],{"type":44,"value":3932},"TE CUDA graphs (attn+mamba+moe), HybridEP",{"type":38,"tag":870,"props":3934,"children":3935},{},[3936,3941,3945,3949],{"type":38,"tag":922,"props":3937,"children":3938},{},[3939],{"type":44,"value":3940},"Nemotron 3 Super",{"type":38,"tag":922,"props":3942,"children":3943},{},[3944],{"type":44,"value":1209},{"type":38,"tag":922,"props":3946,"children":3947},{},[3948],{"type":44,"value":3559},{"type":38,"tag":922,"props":3950,"children":3951},{},[3952],{"type":44,"value":3953},"TE CUDA graphs, EP=64",{"type":38,"tag":870,"props":3955,"children":3956},{},[3957,3962,3966,3971],{"type":38,"tag":922,"props":3958,"children":3959},{},[3960],{"type":44,"value":3961},"NemotronH 56B",{"type":38,"tag":922,"props":3963,"children":3964},{},[3965],{"type":44,"value":1209},{"type":38,"tag":922,"props":3967,"children":3968},{},[3969],{"type":44,"value":3970},"H100, B200, B300",{"type":38,"tag":922,"props":3972,"children":3973},{},[3974],{"type":44,"value":3975},"TP=2–8, TE graphs (mamba+attn)",{"type":38,"tag":250,"props":3977,"children":3979},{"id":3978},"gpt-oss",[3980],{"type":44,"value":3981},"GPT-OSS",{"type":38,"tag":862,"props":3983,"children":3984},{},[3985,4007],{"type":38,"tag":866,"props":3986,"children":3987},{},[3988],{"type":38,"tag":870,"props":3989,"children":3990},{},[3991,3995,3999,4003],{"type":38,"tag":874,"props":3992,"children":3993},{},[3994],{"type":44,"value":3503},{"type":38,"tag":874,"props":3996,"children":3997},{},[3998],{"type":44,"value":1853},{"type":38,"tag":874,"props":4000,"children":4001},{},[4002],{"type":44,"value":3512},{"type":38,"tag":874,"props":4004,"children":4005},{},[4006],{"type":44,"value":3517},{"type":38,"tag":915,"props":4008,"children":4009},{},[4010],{"type":38,"tag":870,"props":4011,"children":4012},{},[4013,4018,4022,4027],{"type":38,"tag":922,"props":4014,"children":4015},{},[4016],{"type":44,"value":4017},"GPT-OSS 120B",{"type":38,"tag":922,"props":4019,"children":4020},{},[4021],{"type":44,"value":1209},{"type":38,"tag":922,"props":4023,"children":4024},{},[4025],{"type":44,"value":4026},"H100, B200, GB200",{"type":38,"tag":922,"props":4028,"children":4029},{},[4030],{"type":44,"value":4031},"EP=64, HybridEP on GB200",{"type":38,"tag":240,"props":4033,"children":4034},{},[],{"type":38,"tag":53,"props":4036,"children":4038},{"id":4037},"recommendation-decision-tree",[4039],{"type":44,"value":4040},"Recommendation Decision Tree",{"type":38,"tag":257,"props":4042,"children":4046},{"className":4043,"code":4045,"language":44,"meta":262},[4044],"language-text","User wants to train a model\n│\n├─ Know the model name?\n│   ├─ Yes → Look up in Library Recipe Index above\n│   │   ├─ Has a recipe for their size + mode? → Use it directly\n│   │   └─ No exact match? → Use closest size, adjust parallelism\n│   └─ No → Ask for model name, size, and HF model ID\n│\n├─ What's the training goal?\n│   ├─ Pretrain → Use *_pretrain_config\n│   ├─ SFT (full fine-tune) → Use *_sft_config\n│   └─ PEFT (LoRA\u002FDoRA) → Use *_peft_config (lowest GPU requirement)\n│\n├─ How many GPUs?\n│   ├─ 1 GPU → Only PEFT recipes work (TP=1, PP=1)\n│   ├─ 8 GPUs (1 node) → Most 8B–16B models, small MoE (EP=8)\n│   ├─ 16–64 GPUs → 70B dense, medium MoE\n│   └─ 128+ GPUs → 405B+, large MoE (DeepSeek V3, Kimi K2)\n│\n├─ Want throughput benchmarks?\n│   ├─ Yes → Use perf recipes (scripts\u002Fperformance\u002F)\n│   │   └─ ⚠️ These run on mock data for upper-bound perf only\n│   └─ No → Use library recipes (scripts\u002Ftraining\u002Frun_recipe.py)\n│\n└─ Long context?\n    ├─ > 8K → Need CP (context parallelism), check *_16k \u002F *_64k \u002F *_128k variants\n    └─ ≤ 8K → Default recipes work\n",[4047],{"type":38,"tag":137,"props":4048,"children":4049},{"__ignoreMap":262},[4050],{"type":44,"value":4045},{"type":38,"tag":240,"props":4052,"children":4053},{},[],{"type":38,"tag":53,"props":4055,"children":4057},{"id":4056},"adjustment-advice-when-recommending",[4058],{"type":44,"value":4059},"Adjustment Advice (When Recommending)",{"type":38,"tag":250,"props":4061,"children":4063},{"id":4062},"parallelism-resizing-rules",[4064],{"type":44,"value":4065},"Parallelism Resizing Rules",{"type":38,"tag":47,"props":4067,"children":4068},{},[4069],{"type":44,"value":4070},"When the user's GPU count differs from the recipe default:",{"type":38,"tag":60,"props":4072,"children":4073},{},[4074,4097,4107,4117,4127,4137,4163],{"type":38,"tag":64,"props":4075,"children":4076},{},[4077,4087,4089,4095],{"type":38,"tag":70,"props":4078,"children":4079},{},[4080,4082],{"type":44,"value":4081},"TP must divide ",{"type":38,"tag":137,"props":4083,"children":4085},{"className":4084},[],[4086],{"type":44,"value":228},{"type":44,"value":4088}," (GQA constraint). E.g. if\n",{"type":38,"tag":137,"props":4090,"children":4092},{"className":4091},[],[4093],{"type":44,"value":4094},"num_key_value_heads=8",{"type":44,"value":4096},", valid TP = {1, 2, 4, 8}.",{"type":38,"tag":64,"props":4098,"children":4099},{},[4100,4105],{"type":38,"tag":70,"props":4101,"children":4102},{},[4103],{"type":44,"value":4104},"TP should stay within a single node",{"type":44,"value":4106}," (NVLink). TP > 8 requires\ninter-node NVLink (e.g., GB200 NVL72).",{"type":38,"tag":64,"props":4108,"children":4109},{},[4110,4115],{"type":38,"tag":70,"props":4111,"children":4112},{},[4113],{"type":44,"value":4114},"PP adds pipeline bubbles.",{"type":44,"value":4116}," Minimize PP; only increase when TP alone can't\nfit the model. Use VP (virtual pipeline) to mitigate bubble overhead.",{"type":38,"tag":64,"props":4118,"children":4119},{},[4120,4125],{"type":38,"tag":70,"props":4121,"children":4122},{},[4123],{"type":44,"value":4124},"EP doesn't reduce dense-layer memory.",{"type":44,"value":4126}," Only expert parameters shard with\nEP. Shared attention\u002Fembeddings are replicated. For \"OOM with MoE\", increase\nEP first, not TP.",{"type":38,"tag":64,"props":4128,"children":4129},{},[4130,4135],{"type":38,"tag":70,"props":4131,"children":4132},{},[4133],{"type":44,"value":4134},"SP should be True whenever TP > 1.",{"type":44,"value":4136}," It eliminates redundant activation\ncopies and is essentially free.",{"type":38,"tag":64,"props":4138,"children":4139},{},[4140,4145,4147,4153,4155,4161],{"type":38,"tag":70,"props":4141,"children":4142},{},[4143],{"type":44,"value":4144},"CP requires all-to-all or ring attention.",{"type":44,"value":4146}," Check ",{"type":38,"tag":137,"props":4148,"children":4150},{"className":4149},[],[4151],{"type":44,"value":4152},"cp_comm_type",{"type":44,"value":4154},". For\nGQA models, ",{"type":38,"tag":137,"props":4156,"children":4158},{"className":4157},[],[4159],{"type":44,"value":4160},"a2a+p2p",{"type":44,"value":4162}," hierarchical CP allows CP > num_kv_heads.",{"type":38,"tag":64,"props":4164,"children":4165},{},[4166,4171],{"type":38,"tag":70,"props":4167,"children":4168},{},[4169],{"type":44,"value":4170},"world_size = DP × TP × PP × CP × EP.",{"type":44,"value":4172}," DP is implicit. Make sure the\nproduct of explicit parallelisms divides your total GPU count.",{"type":38,"tag":250,"props":4174,"children":4176},{"id":4175},"batch-size-tuning",[4177],{"type":44,"value":4178},"Batch Size Tuning",{"type":38,"tag":732,"props":4180,"children":4181},{},[4182,4194,4212],{"type":38,"tag":64,"props":4183,"children":4184},{},[4185,4187,4192],{"type":44,"value":4186},"Start with the recipe's ",{"type":38,"tag":137,"props":4188,"children":4190},{"className":4189},[],[4191],{"type":44,"value":236},{"type":44,"value":4193},". If OOM, reduce to 1.",{"type":38,"tag":64,"props":4195,"children":4196},{},[4197,4203,4205,4211],{"type":38,"tag":137,"props":4198,"children":4200},{"className":4199},[],[4201],{"type":44,"value":4202},"global_batch_size",{"type":44,"value":4204}," determines learning dynamics. Scale with DP:\n",{"type":38,"tag":137,"props":4206,"children":4208},{"className":4207},[],[4209],{"type":44,"value":4210},"GBS = micro_batch_size × DP × gradient_accumulation_steps",{"type":44,"value":152},{"type":38,"tag":64,"props":4213,"children":4214},{},[4215,4217,4223],{"type":44,"value":4216},"For MoE, ",{"type":38,"tag":137,"props":4218,"children":4220},{"className":4219},[],[4221],{"type":44,"value":4222},"micro_batch_size=1",{"type":44,"value":4224}," is typical at scale.",{"type":38,"tag":250,"props":4226,"children":4228},{"id":4227},"common-pitfalls-to-warn-about",[4229],{"type":44,"value":4230},"Common Pitfalls to Warn About",{"type":38,"tag":862,"props":4232,"children":4233},{},[4234,4255],{"type":38,"tag":866,"props":4235,"children":4236},{},[4237],{"type":38,"tag":870,"props":4238,"children":4239},{},[4240,4245,4250],{"type":38,"tag":874,"props":4241,"children":4242},{},[4243],{"type":44,"value":4244},"Pitfall",{"type":38,"tag":874,"props":4246,"children":4247},{},[4248],{"type":44,"value":4249},"Symptom",{"type":38,"tag":874,"props":4251,"children":4252},{},[4253],{"type":44,"value":4254},"Fix",{"type":38,"tag":915,"props":4256,"children":4257},{},[4258,4276,4300,4318,4344,4369,4406,4424,4465,4498,4516],{"type":38,"tag":870,"props":4259,"children":4260},{},[4261,4266,4271],{"type":38,"tag":922,"props":4262,"children":4263},{},[4264],{"type":44,"value":4265},"TP > num_kv_heads",{"type":38,"tag":922,"props":4267,"children":4268},{},[4269],{"type":44,"value":4270},"Crash: \"TP must divide num_query_groups\"",{"type":38,"tag":922,"props":4272,"children":4273},{},[4274],{"type":44,"value":4275},"Reduce TP to a divisor of num_kv_heads",{"type":38,"tag":870,"props":4277,"children":4278},{},[4279,4284,4289],{"type":38,"tag":922,"props":4280,"children":4281},{},[4282],{"type":44,"value":4283},"PP without VP",{"type":38,"tag":922,"props":4285,"children":4286},{},[4287],{"type":44,"value":4288},"Poor throughput (large bubble)",{"type":38,"tag":922,"props":4290,"children":4291},{},[4292,4294],{"type":44,"value":4293},"Set ",{"type":38,"tag":137,"props":4295,"children":4297},{"className":4296},[],[4298],{"type":44,"value":4299},"virtual_pipeline_model_parallel_size",{"type":38,"tag":870,"props":4301,"children":4302},{},[4303,4308,4313],{"type":38,"tag":922,"props":4304,"children":4305},{},[4306],{"type":44,"value":4307},"EP too low for large MoE",{"type":38,"tag":922,"props":4309,"children":4310},{},[4311],{"type":44,"value":4312},"OOM on expert params",{"type":38,"tag":922,"props":4314,"children":4315},{},[4316],{"type":44,"value":4317},"Increase EP; each expert lives on EP\u002Fnum_experts ranks",{"type":38,"tag":870,"props":4319,"children":4320},{},[4321,4326,4331],{"type":38,"tag":922,"props":4322,"children":4323},{},[4324],{"type":44,"value":4325},"CUDA graphs + packed sequences",{"type":38,"tag":922,"props":4327,"children":4328},{},[4329],{"type":44,"value":4330},"Assert: \"CUDA graph accepts only Tensor inputs\"",{"type":38,"tag":922,"props":4332,"children":4333},{},[4334,4336,4342],{"type":44,"value":4335},"Disable packing or use ",{"type":38,"tag":137,"props":4337,"children":4339},{"className":4338},[],[4340],{"type":44,"value":4341},"local",{"type":44,"value":4343}," full-iteration graphs",{"type":38,"tag":870,"props":4345,"children":4346},{},[4347,4352,4357],{"type":38,"tag":922,"props":4348,"children":4349},{},[4350],{"type":44,"value":4351},"CUDA graphs + full recompute",{"type":38,"tag":922,"props":4353,"children":4354},{},[4355],{"type":44,"value":4356},"Assert: \"full recompute only with full iteration CUDA graph\"",{"type":38,"tag":922,"props":4358,"children":4359},{},[4360,4362,4367],{"type":44,"value":4361},"Disable recompute or switch to ",{"type":38,"tag":137,"props":4363,"children":4365},{"className":4364},[],[4366],{"type":44,"value":4341},{"type":44,"value":4368}," impl",{"type":38,"tag":870,"props":4370,"children":4371},{},[4372,4383,4388],{"type":38,"tag":922,"props":4373,"children":4374},{},[4375,4381],{"type":38,"tag":137,"props":4376,"children":4378},{"className":4377},[],[4379],{"type":44,"value":4380},"use_te_rng_tracker",{"type":44,"value":4382}," not set",{"type":38,"tag":922,"props":4384,"children":4385},{},[4386],{"type":44,"value":4387},"Assert on provider init when CUDA graphs enabled",{"type":38,"tag":922,"props":4389,"children":4390},{},[4391,4392,4398,4400],{"type":44,"value":4293},{"type":38,"tag":137,"props":4393,"children":4395},{"className":4394},[],[4396],{"type":44,"value":4397},"cfg.model.use_te_rng_tracker = True",{"type":44,"value":4399}," and ",{"type":38,"tag":137,"props":4401,"children":4403},{"className":4402},[],[4404],{"type":44,"value":4405},"cfg.rng.te_rng_tracker = True",{"type":38,"tag":870,"props":4407,"children":4408},{},[4409,4414,4419],{"type":38,"tag":922,"props":4410,"children":4411},{},[4412],{"type":44,"value":4413},"FSDP + TP > 1 on H100",{"type":38,"tag":922,"props":4415,"children":4416},{},[4417],{"type":44,"value":4418},"Possible comm bottleneck",{"type":38,"tag":922,"props":4420,"children":4421},{},[4422],{"type":44,"value":4423},"Prefer FSDP with TP=1 or TP=2 on H100; FSDP shines on GB\u002FB-series",{"type":38,"tag":870,"props":4425,"children":4426},{},[4427,4432,4437],{"type":38,"tag":922,"props":4428,"children":4429},{},[4430],{"type":44,"value":4431},"Long context without CP",{"type":38,"tag":922,"props":4433,"children":4434},{},[4435],{"type":44,"value":4436},"OOM on activations",{"type":38,"tag":922,"props":4438,"children":4439},{},[4440,4442,4448,4449,4455,4457,4463],{"type":44,"value":4441},"Add CP=2\u002F4\u002F8; use ",{"type":38,"tag":137,"props":4443,"children":4445},{"className":4444},[],[4446],{"type":44,"value":4447},"*_16k",{"type":44,"value":76},{"type":38,"tag":137,"props":4450,"children":4452},{"className":4451},[],[4453],{"type":44,"value":4454},"*_64k",{"type":44,"value":4456},", or ",{"type":38,"tag":137,"props":4458,"children":4460},{"className":4459},[],[4461],{"type":44,"value":4462},"*_128k",{"type":44,"value":4464}," recipe variants",{"type":38,"tag":870,"props":4466,"children":4467},{},[4468,4481,4486],{"type":38,"tag":922,"props":4469,"children":4470},{},[4471,4473,4479],{"type":44,"value":4472},"MoE ",{"type":38,"tag":137,"props":4474,"children":4476},{"className":4475},[],[4477],{"type":44,"value":4478},"overlap_grad_reduce",{"type":44,"value":4480}," on H100",{"type":38,"tag":922,"props":4482,"children":4483},{},[4484],{"type":44,"value":4485},"May hurt perf (False in many H100 presets)",{"type":38,"tag":922,"props":4487,"children":4488},{},[4489,4490,4496],{"type":44,"value":4293},{"type":38,"tag":137,"props":4491,"children":4493},{"className":4492},[],[4494],{"type":44,"value":4495},"overlap_grad_reduce=False",{"type":44,"value":4497}," for MoE on H100",{"type":38,"tag":870,"props":4499,"children":4500},{},[4501,4506,4511],{"type":38,"tag":922,"props":4502,"children":4503},{},[4504],{"type":44,"value":4505},"VLM SFT missing image data",{"type":38,"tag":922,"props":4507,"children":4508},{},[4509],{"type":44,"value":4510},"Runs but produces garbage",{"type":38,"tag":922,"props":4512,"children":4513},{},[4514],{"type":44,"value":4515},"Provide actual multimodal dataset or use mock VLM data",{"type":38,"tag":870,"props":4517,"children":4518},{},[4519,4524,4529],{"type":38,"tag":922,"props":4520,"children":4521},{},[4522],{"type":44,"value":4523},"Qwen35-VL MoE FSDP",{"type":38,"tag":922,"props":4525,"children":4526},{},[4527],{"type":44,"value":4528},"Tested on Blackwell only",{"type":38,"tag":922,"props":4530,"children":4531},{},[4532],{"type":44,"value":4533},"May not work on H100; validate first",{"type":38,"tag":250,"props":4535,"children":4537},{"id":4536},"recipe-override-examples",[4538],{"type":44,"value":4539},"Recipe Override Examples",{"type":38,"tag":257,"props":4541,"children":4543},{"className":259,"code":4542,"language":261,"meta":262,"style":262},"# Scale Llama3 8B from 2 GPUs to 8 GPUs (increase DP)\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe llama3_8b_pretrain_config \\\n    --dataset llm-pretrain-mock\n\n# Reduce parallelism for Qwen3-MoE 30B to fit on 4 GPUs\nuv run python -m torch.distributed.run --nproc_per_node=4 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe qwen3_30b_a3b_sft_config \\\n    --dataset llm-finetune \\\n    'model.expert_model_parallel_size=4'\n\n# Add long context to an existing recipe\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe llama3_8b_pretrain_config \\\n    --dataset llm-pretrain-mock \\\n    'model.seq_length=32768' \\\n    'model.context_parallel_size=4'\n\n# Enable CUDA graphs on any recipe\nuv run python -m torch.distributed.run --nproc_per_node=8 scripts\u002Ftraining\u002Frun_recipe.py \\\n    --recipe qwen3_30b_a3b_pretrain_config \\\n    --dataset llm-pretrain-mock \\\n    'model.cuda_graph_impl=transformer_engine' \\\n    'model.cuda_graph_scope=[attn,moe_router,moe_preprocess]' \\\n    'model.use_te_rng_tracker=True' \\\n    'rng.te_rng_tracker=True'\n",[4544],{"type":38,"tag":137,"props":4545,"children":4546},{"__ignoreMap":262},[4547,4555,4590,4605,4616,4623,4631,4667,4683,4699,4715,4722,4730,4765,4780,4795,4815,4832,4840,4849,4885,4902,4918,4939,4960,4981],{"type":38,"tag":268,"props":4548,"children":4549},{"class":270,"line":271},[4550],{"type":38,"tag":268,"props":4551,"children":4552},{"style":275},[4553],{"type":44,"value":4554},"# Scale Llama3 8B from 2 GPUs to 8 GPUs (increase DP)\n",{"type":38,"tag":268,"props":4556,"children":4557},{"class":270,"line":281},[4558,4562,4566,4570,4574,4578,4582,4586],{"type":38,"tag":268,"props":4559,"children":4560},{"style":285},[4561],{"type":44,"value":288},{"type":38,"tag":268,"props":4563,"children":4564},{"style":291},[4565],{"type":44,"value":294},{"type":38,"tag":268,"props":4567,"children":4568},{"style":291},[4569],{"type":44,"value":299},{"type":38,"tag":268,"props":4571,"children":4572},{"style":291},[4573],{"type":44,"value":304},{"type":38,"tag":268,"props":4575,"children":4576},{"style":291},[4577],{"type":44,"value":309},{"type":38,"tag":268,"props":4579,"children":4580},{"style":291},[4581],{"type":44,"value":314},{"type":38,"tag":268,"props":4583,"children":4584},{"style":291},[4585],{"type":44,"value":319},{"type":38,"tag":268,"props":4587,"children":4588},{"style":322},[4589],{"type":44,"value":325},{"type":38,"tag":268,"props":4591,"children":4592},{"class":270,"line":328},[4593,4597,4601],{"type":38,"tag":268,"props":4594,"children":4595},{"style":291},[4596],{"type":44,"value":334},{"type":38,"tag":268,"props":4598,"children":4599},{"style":291},[4600],{"type":44,"value":535},{"type":38,"tag":268,"props":4602,"children":4603},{"style":322},[4604],{"type":44,"value":325},{"type":38,"tag":268,"props":4606,"children":4607},{"class":270,"line":362},[4608,4612],{"type":38,"tag":268,"props":4609,"children":4610},{"style":291},[4611],{"type":44,"value":368},{"type":38,"tag":268,"props":4613,"children":4614},{"style":291},[4615],{"type":44,"value":373},{"type":38,"tag":268,"props":4617,"children":4618},{"class":270,"line":376},[4619],{"type":38,"tag":268,"props":4620,"children":4621},{"emptyLinePlaceholder":380},[4622],{"type":44,"value":383},{"type":38,"tag":268,"props":4624,"children":4625},{"class":270,"line":386},[4626],{"type":38,"tag":268,"props":4627,"children":4628},{"style":275},[4629],{"type":44,"value":4630},"# Reduce parallelism for Qwen3-MoE 30B to fit on 4 GPUs\n",{"type":38,"tag":268,"props":4632,"children":4633},{"class":270,"line":395},[4634,4638,4642,4646,4650,4654,4659,4663],{"type":38,"tag":268,"props":4635,"children":4636},{"style":285},[4637],{"type":44,"value":288},{"type":38,"tag":268,"props":4639,"children":4640},{"style":291},[4641],{"type":44,"value":294},{"type":38,"tag":268,"props":4643,"children":4644},{"style":291},[4645],{"type":44,"value":299},{"type":38,"tag":268,"props":4647,"children":4648},{"style":291},[4649],{"type":44,"value":304},{"type":38,"tag":268,"props":4651,"children":4652},{"style":291},[4653],{"type":44,"value":309},{"type":38,"tag":268,"props":4655,"children":4656},{"style":291},[4657],{"type":44,"value":4658}," --nproc_per_node=4",{"type":38,"tag":268,"props":4660,"children":4661},{"style":291},[4662],{"type":44,"value":319},{"type":38,"tag":268,"props":4664,"children":4665},{"style":322},[4666],{"type":44,"value":325},{"type":38,"tag":268,"props":4668,"children":4669},{"class":270,"line":431},[4670,4674,4679],{"type":38,"tag":268,"props":4671,"children":4672},{"style":291},[4673],{"type":44,"value":334},{"type":38,"tag":268,"props":4675,"children":4676},{"style":291},[4677],{"type":44,"value":4678}," qwen3_30b_a3b_sft_config",{"type":38,"tag":268,"props":4680,"children":4681},{"style":322},[4682],{"type":44,"value":325},{"type":38,"tag":268,"props":4684,"children":4685},{"class":270,"line":459},[4686,4690,4695],{"type":38,"tag":268,"props":4687,"children":4688},{"style":291},[4689],{"type":44,"value":368},{"type":38,"tag":268,"props":4691,"children":4692},{"style":291},[4693],{"type":44,"value":4694}," llm-finetune",{"type":38,"tag":268,"props":4696,"children":4697},{"style":322},[4698],{"type":44,"value":325},{"type":38,"tag":268,"props":4700,"children":4701},{"class":270,"line":472},[4702,4706,4711],{"type":38,"tag":268,"props":4703,"children":4704},{"style":337},[4705],{"type":44,"value":565},{"type":38,"tag":268,"props":4707,"children":4708},{"style":291},[4709],{"type":44,"value":4710},"model.expert_model_parallel_size=4",{"type":38,"tag":268,"props":4712,"children":4713},{"style":337},[4714],{"type":44,"value":597},{"type":38,"tag":268,"props":4716,"children":4717},{"class":270,"line":480},[4718],{"type":38,"tag":268,"props":4719,"children":4720},{"emptyLinePlaceholder":380},[4721],{"type":44,"value":383},{"type":38,"tag":268,"props":4723,"children":4724},{"class":270,"line":489},[4725],{"type":38,"tag":268,"props":4726,"children":4727},{"style":275},[4728],{"type":44,"value":4729},"# Add long context to an existing recipe\n",{"type":38,"tag":268,"props":4731,"children":4732},{"class":270,"line":525},[4733,4737,4741,4745,4749,4753,4757,4761],{"type":38,"tag":268,"props":4734,"children":4735},{"style":285},[4736],{"type":44,"value":288},{"type":38,"tag":268,"props":4738,"children":4739},{"style":291},[4740],{"type":44,"value":294},{"type":38,"tag":268,"props":4742,"children":4743},{"style":291},[4744],{"type":44,"value":299},{"type":38,"tag":268,"props":4746,"children":4747},{"style":291},[4748],{"type":44,"value":304},{"type":38,"tag":268,"props":4750,"children":4751},{"style":291},[4752],{"type":44,"value":309},{"type":38,"tag":268,"props":4754,"children":4755},{"style":291},[4756],{"type":44,"value":314},{"type":38,"tag":268,"props":4758,"children":4759},{"style":291},[4760],{"type":44,"value":319},{"type":38,"tag":268,"props":4762,"children":4763},{"style":322},[4764],{"type":44,"value":325},{"type":38,"tag":268,"props":4766,"children":4767},{"class":270,"line":542},[4768,4772,4776],{"type":38,"tag":268,"props":4769,"children":4770},{"style":291},[4771],{"type":44,"value":334},{"type":38,"tag":268,"props":4773,"children":4774},{"style":291},[4775],{"type":44,"value":535},{"type":38,"tag":268,"props":4777,"children":4778},{"style":322},[4779],{"type":44,"value":325},{"type":38,"tag":268,"props":4781,"children":4782},{"class":270,"line":559},[4783,4787,4791],{"type":38,"tag":268,"props":4784,"children":4785},{"style":291},[4786],{"type":44,"value":368},{"type":38,"tag":268,"props":4788,"children":4789},{"style":291},[4790],{"type":44,"value":552},{"type":38,"tag":268,"props":4792,"children":4793},{"style":322},[4794],{"type":44,"value":325},{"type":38,"tag":268,"props":4796,"children":4797},{"class":270,"line":582},[4798,4802,4807,4811],{"type":38,"tag":268,"props":4799,"children":4800},{"style":337},[4801],{"type":44,"value":565},{"type":38,"tag":268,"props":4803,"children":4804},{"style":291},[4805],{"type":44,"value":4806},"model.seq_length=32768",{"type":38,"tag":268,"props":4808,"children":4809},{"style":337},[4810],{"type":44,"value":575},{"type":38,"tag":268,"props":4812,"children":4813},{"style":322},[4814],{"type":44,"value":325},{"type":38,"tag":268,"props":4816,"children":4818},{"class":270,"line":4817},17,[4819,4823,4828],{"type":38,"tag":268,"props":4820,"children":4821},{"style":337},[4822],{"type":44,"value":565},{"type":38,"tag":268,"props":4824,"children":4825},{"style":291},[4826],{"type":44,"value":4827},"model.context_parallel_size=4",{"type":38,"tag":268,"props":4829,"children":4830},{"style":337},[4831],{"type":44,"value":597},{"type":38,"tag":268,"props":4833,"children":4835},{"class":270,"line":4834},18,[4836],{"type":38,"tag":268,"props":4837,"children":4838},{"emptyLinePlaceholder":380},[4839],{"type":44,"value":383},{"type":38,"tag":268,"props":4841,"children":4843},{"class":270,"line":4842},19,[4844],{"type":38,"tag":268,"props":4845,"children":4846},{"style":275},[4847],{"type":44,"value":4848},"# Enable CUDA graphs on any recipe\n",{"type":38,"tag":268,"props":4850,"children":4852},{"class":270,"line":4851},20,[4853,4857,4861,4865,4869,4873,4877,4881],{"type":38,"tag":268,"props":4854,"children":4855},{"style":285},[4856],{"type":44,"value":288},{"type":38,"tag":268,"props":4858,"children":4859},{"style":291},[4860],{"type":44,"value":294},{"type":38,"tag":268,"props":4862,"children":4863},{"style":291},[4864],{"type":44,"value":299},{"type":38,"tag":268,"props":4866,"children":4867},{"style":291},[4868],{"type":44,"value":304},{"type":38,"tag":268,"props":4870,"children":4871},{"style":291},[4872],{"type":44,"value":309},{"type":38,"tag":268,"props":4874,"children":4875},{"style":291},[4876],{"type":44,"value":314},{"type":38,"tag":268,"props":4878,"children":4879},{"style":291},[4880],{"type":44,"value":319},{"type":38,"tag":268,"props":4882,"children":4883},{"style":322},[4884],{"type":44,"value":325},{"type":38,"tag":268,"props":4886,"children":4888},{"class":270,"line":4887},21,[4889,4893,4898],{"type":38,"tag":268,"props":4890,"children":4891},{"style":291},[4892],{"type":44,"value":334},{"type":38,"tag":268,"props":4894,"children":4895},{"style":291},[4896],{"type":44,"value":4897}," qwen3_30b_a3b_pretrain_config",{"type":38,"tag":268,"props":4899,"children":4900},{"style":322},[4901],{"type":44,"value":325},{"type":38,"tag":268,"props":4903,"children":4905},{"class":270,"line":4904},22,[4906,4910,4914],{"type":38,"tag":268,"props":4907,"children":4908},{"style":291},[4909],{"type":44,"value":368},{"type":38,"tag":268,"props":4911,"children":4912},{"style":291},[4913],{"type":44,"value":552},{"type":38,"tag":268,"props":4915,"children":4916},{"style":322},[4917],{"type":44,"value":325},{"type":38,"tag":268,"props":4919,"children":4921},{"class":270,"line":4920},23,[4922,4926,4931,4935],{"type":38,"tag":268,"props":4923,"children":4924},{"style":337},[4925],{"type":44,"value":565},{"type":38,"tag":268,"props":4927,"children":4928},{"style":291},[4929],{"type":44,"value":4930},"model.cuda_graph_impl=transformer_engine",{"type":38,"tag":268,"props":4932,"children":4933},{"style":337},[4934],{"type":44,"value":575},{"type":38,"tag":268,"props":4936,"children":4937},{"style":322},[4938],{"type":44,"value":325},{"type":38,"tag":268,"props":4940,"children":4942},{"class":270,"line":4941},24,[4943,4947,4952,4956],{"type":38,"tag":268,"props":4944,"children":4945},{"style":337},[4946],{"type":44,"value":565},{"type":38,"tag":268,"props":4948,"children":4949},{"style":291},[4950],{"type":44,"value":4951},"model.cuda_graph_scope=[attn,moe_router,moe_preprocess]",{"type":38,"tag":268,"props":4953,"children":4954},{"style":337},[4955],{"type":44,"value":575},{"type":38,"tag":268,"props":4957,"children":4958},{"style":322},[4959],{"type":44,"value":325},{"type":38,"tag":268,"props":4961,"children":4963},{"class":270,"line":4962},25,[4964,4968,4973,4977],{"type":38,"tag":268,"props":4965,"children":4966},{"style":337},[4967],{"type":44,"value":565},{"type":38,"tag":268,"props":4969,"children":4970},{"style":291},[4971],{"type":44,"value":4972},"model.use_te_rng_tracker=True",{"type":38,"tag":268,"props":4974,"children":4975},{"style":337},[4976],{"type":44,"value":575},{"type":38,"tag":268,"props":4978,"children":4979},{"style":322},[4980],{"type":44,"value":325},{"type":38,"tag":268,"props":4982,"children":4984},{"class":270,"line":4983},26,[4985,4989,4994],{"type":38,"tag":268,"props":4986,"children":4987},{"style":337},[4988],{"type":44,"value":565},{"type":38,"tag":268,"props":4990,"children":4991},{"style":291},[4992],{"type":44,"value":4993},"rng.te_rng_tracker=True",{"type":38,"tag":268,"props":4995,"children":4996},{"style":337},[4997],{"type":44,"value":597},{"type":38,"tag":240,"props":4999,"children":5000},{},[],{"type":38,"tag":53,"props":5002,"children":5004},{"id":5003},"quick-reference-which-recipe-for-my-situation",[5005],{"type":44,"value":5006},"Quick Reference: Which Recipe for My Situation?",{"type":38,"tag":862,"props":5008,"children":5009},{},[5010,5031],{"type":38,"tag":866,"props":5011,"children":5012},{},[5013],{"type":38,"tag":870,"props":5014,"children":5015},{},[5016,5021,5026],{"type":38,"tag":874,"props":5017,"children":5018},{},[5019],{"type":44,"value":5020},"I want to...",{"type":38,"tag":874,"props":5022,"children":5023},{},[5024],{"type":44,"value":5025},"Start with",{"type":38,"tag":874,"props":5027,"children":5028},{},[5029],{"type":44,"value":5030},"GPUs needed",{"type":38,"tag":915,"props":5032,"children":5033},{},[5034,5056,5084,5111,5132,5152,5173,5197,5220,5249],{"type":38,"tag":870,"props":5035,"children":5036},{},[5037,5042,5052],{"type":38,"tag":922,"props":5038,"children":5039},{},[5040],{"type":44,"value":5041},"Try Bridge for the first time",{"type":38,"tag":922,"props":5043,"children":5044},{},[5045,5050],{"type":38,"tag":137,"props":5046,"children":5048},{"className":5047},[],[5049],{"type":44,"value":180},{"type":44,"value":5051}," + mock data",{"type":38,"tag":922,"props":5053,"children":5054},{},[5055],{"type":44,"value":940},{"type":38,"tag":870,"props":5057,"children":5058},{},[5059,5064,5080],{"type":38,"tag":922,"props":5060,"children":5061},{},[5062],{"type":44,"value":5063},"Fine-tune a 7-8B model",{"type":38,"tag":922,"props":5065,"children":5066},{},[5067,5072,5074],{"type":38,"tag":137,"props":5068,"children":5070},{"className":5069},[],[5071],{"type":44,"value":180},{"type":44,"value":5073}," or ",{"type":38,"tag":137,"props":5075,"children":5077},{"className":5076},[],[5078],{"type":44,"value":5079},"qwen3_8b_sft_config",{"type":38,"tag":922,"props":5081,"children":5082},{},[5083],{"type":44,"value":2660},{"type":38,"tag":870,"props":5085,"children":5086},{},[5087,5092,5107],{"type":38,"tag":922,"props":5088,"children":5089},{},[5090],{"type":44,"value":5091},"LoRA on 1 GPU",{"type":38,"tag":922,"props":5093,"children":5094},{},[5095,5100,5101],{"type":38,"tag":137,"props":5096,"children":5098},{"className":5097},[],[5099],{"type":44,"value":1426},{"type":44,"value":5073},{"type":38,"tag":137,"props":5102,"children":5104},{"className":5103},[],[5105],{"type":44,"value":5106},"qwen3_8b_peft_config",{"type":38,"tag":922,"props":5108,"children":5109},{},[5110],{"type":44,"value":945},{"type":38,"tag":870,"props":5112,"children":5113},{},[5114,5119,5127],{"type":38,"tag":922,"props":5115,"children":5116},{},[5117],{"type":44,"value":5118},"Pretrain a dense 70B",{"type":38,"tag":922,"props":5120,"children":5121},{},[5122],{"type":38,"tag":137,"props":5123,"children":5125},{"className":5124},[],[5126],{"type":44,"value":1143},{"type":38,"tag":922,"props":5128,"children":5129},{},[5130],{"type":44,"value":5131},"32–64",{"type":38,"tag":870,"props":5133,"children":5134},{},[5135,5140,5148],{"type":38,"tag":922,"props":5136,"children":5137},{},[5138],{"type":44,"value":5139},"Train a small MoE",{"type":38,"tag":922,"props":5141,"children":5142},{},[5143],{"type":38,"tag":137,"props":5144,"children":5146},{"className":5145},[],[5147],{"type":44,"value":1868},{"type":38,"tag":922,"props":5149,"children":5150},{},[5151],{"type":44,"value":1084},{"type":38,"tag":870,"props":5153,"children":5154},{},[5155,5160,5168],{"type":38,"tag":922,"props":5156,"children":5157},{},[5158],{"type":44,"value":5159},"Train a large MoE (235B+)",{"type":38,"tag":922,"props":5161,"children":5162},{},[5163],{"type":38,"tag":137,"props":5164,"children":5166},{"className":5165},[],[5167],{"type":44,"value":1976},{"type":38,"tag":922,"props":5169,"children":5170},{},[5171],{"type":44,"value":5172},"256–512",{"type":38,"tag":870,"props":5174,"children":5175},{},[5176,5181,5192],{"type":38,"tag":922,"props":5177,"children":5178},{},[5179],{"type":44,"value":5180},"Benchmark throughput",{"type":38,"tag":922,"props":5182,"children":5183},{},[5184,5186],{"type":44,"value":5185},"Perf recipes via ",{"type":38,"tag":137,"props":5187,"children":5189},{"className":5188},[],[5190],{"type":44,"value":5191},"run_script.py",{"type":38,"tag":922,"props":5193,"children":5194},{},[5195],{"type":44,"value":5196},"Varies",{"type":38,"tag":870,"props":5198,"children":5199},{},[5200,5205,5215],{"type":38,"tag":922,"props":5201,"children":5202},{},[5203],{"type":44,"value":5204},"Long-context training",{"type":38,"tag":922,"props":5206,"children":5207},{},[5208,5213],{"type":38,"tag":137,"props":5209,"children":5211},{"className":5210},[],[5212],{"type":44,"value":1101},{"type":44,"value":5214}," or add CP override",{"type":38,"tag":922,"props":5216,"children":5217},{},[5218],{"type":44,"value":5219},"16+",{"type":38,"tag":870,"props":5221,"children":5222},{},[5223,5228,5244],{"type":38,"tag":922,"props":5224,"children":5225},{},[5226],{"type":44,"value":5227},"VLM fine-tuning",{"type":38,"tag":922,"props":5229,"children":5230},{},[5231,5237,5238],{"type":38,"tag":137,"props":5232,"children":5234},{"className":5233},[],[5235],{"type":44,"value":5236},"qwen3_vl_8b_sft_config",{"type":44,"value":5073},{"type":38,"tag":137,"props":5239,"children":5241},{"className":5240},[],[5242],{"type":44,"value":5243},"gemma3_vl_*_sft_config",{"type":38,"tag":922,"props":5245,"children":5246},{},[5247],{"type":44,"value":5248},"4–8",{"type":38,"tag":870,"props":5250,"children":5251},{},[5252,5257,5273],{"type":38,"tag":922,"props":5253,"children":5254},{},[5255],{"type":44,"value":5256},"Diffusion training",{"type":38,"tag":922,"props":5258,"children":5259},{},[5260,5266,5267],{"type":38,"tag":137,"props":5261,"children":5263},{"className":5262},[],[5264],{"type":44,"value":5265},"wan_1_3B_pretrain_config",{"type":44,"value":5073},{"type":38,"tag":137,"props":5268,"children":5270},{"className":5269},[],[5271],{"type":44,"value":5272},"flux_12b_pretrain_config",{"type":38,"tag":922,"props":5274,"children":5275},{},[5276],{"type":44,"value":1084},{"type":38,"tag":240,"props":5278,"children":5279},{},[],{"type":38,"tag":53,"props":5281,"children":5283},{"id":5282},"code-anchors",[5284],{"type":44,"value":5285},"Code Anchors",{"type":38,"tag":862,"props":5287,"children":5288},{},[5289,5305],{"type":38,"tag":866,"props":5290,"children":5291},{},[5292],{"type":38,"tag":870,"props":5293,"children":5294},{},[5295,5300],{"type":38,"tag":874,"props":5296,"children":5297},{},[5298],{"type":44,"value":5299},"What",{"type":38,"tag":874,"props":5301,"children":5302},{},[5303],{"type":44,"value":5304},"Path",{"type":38,"tag":915,"props":5306,"children":5307},{},[5308,5324,5349,5366,5382,5398,5415,5431],{"type":38,"tag":870,"props":5309,"children":5310},{},[5311,5316],{"type":38,"tag":922,"props":5312,"children":5313},{},[5314],{"type":44,"value":5315},"Library recipes root",{"type":38,"tag":922,"props":5317,"children":5318},{},[5319],{"type":38,"tag":137,"props":5320,"children":5322},{"className":5321},[],[5323],{"type":44,"value":142},{"type":38,"tag":870,"props":5325,"children":5326},{},[5327,5340],{"type":38,"tag":922,"props":5328,"children":5329},{},[5330,5332,5338],{"type":44,"value":5331},"Recipe ",{"type":38,"tag":137,"props":5333,"children":5335},{"className":5334},[],[5336],{"type":44,"value":5337},"__init__.py",{"type":44,"value":5339}," (all exports)",{"type":38,"tag":922,"props":5341,"children":5342},{},[5343],{"type":38,"tag":137,"props":5344,"children":5346},{"className":5345},[],[5347],{"type":44,"value":5348},"src\u002Fmegatron\u002Fbridge\u002Frecipes\u002F__init__.py",{"type":38,"tag":870,"props":5350,"children":5351},{},[5352,5357],{"type":38,"tag":922,"props":5353,"children":5354},{},[5355],{"type":44,"value":5356},"Common recipe helpers",{"type":38,"tag":922,"props":5358,"children":5359},{},[5360],{"type":38,"tag":137,"props":5361,"children":5363},{"className":5362},[],[5364],{"type":44,"value":5365},"src\u002Fmegatron\u002Fbridge\u002Frecipes\u002Fcommon.py",{"type":38,"tag":870,"props":5367,"children":5368},{},[5369,5374],{"type":38,"tag":922,"props":5370,"children":5371},{},[5372],{"type":44,"value":5373},"Training entry point",{"type":38,"tag":922,"props":5375,"children":5376},{},[5377],{"type":38,"tag":137,"props":5378,"children":5380},{"className":5379},[],[5381],{"type":44,"value":150},{"type":38,"tag":870,"props":5383,"children":5384},{},[5385,5390],{"type":38,"tag":922,"props":5386,"children":5387},{},[5388],{"type":44,"value":5389},"Perf recipes root",{"type":38,"tag":922,"props":5391,"children":5392},{},[5393],{"type":38,"tag":137,"props":5394,"children":5396},{"className":5395},[],[5397],{"type":44,"value":3439},{"type":38,"tag":870,"props":5399,"children":5400},{},[5401,5406],{"type":38,"tag":922,"props":5402,"children":5403},{},[5404],{"type":44,"value":5405},"Perf entry point",{"type":38,"tag":922,"props":5407,"children":5408},{},[5409],{"type":38,"tag":137,"props":5410,"children":5412},{"className":5411},[],[5413],{"type":44,"value":5414},"scripts\u002Fperformance\u002Frun_script.py",{"type":38,"tag":870,"props":5416,"children":5417},{},[5418,5423],{"type":38,"tag":922,"props":5419,"children":5420},{},[5421],{"type":44,"value":5422},"Perf recipe helpers",{"type":38,"tag":922,"props":5424,"children":5425},{},[5426],{"type":38,"tag":137,"props":5427,"children":5429},{"className":5428},[],[5430],{"type":44,"value":779},{"type":38,"tag":870,"props":5432,"children":5433},{},[5434,5439],{"type":38,"tag":922,"props":5435,"children":5436},{},[5437],{"type":44,"value":5438},"Perf overrides (benchmark defaults)",{"type":38,"tag":922,"props":5440,"children":5441},{},[5442],{"type":38,"tag":137,"props":5443,"children":5445},{"className":5444},[],[5446],{"type":44,"value":5447},"scripts\u002Fperformance\u002Futils\u002Foverrides.py",{"type":38,"tag":5449,"props":5450,"children":5451},"style",{},[5452],{"type":44,"value":5453},"html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"items":5455,"total":5559},[5456,5473,5487,5501,5513,5530,5545],{"slug":5457,"name":5457,"fn":5458,"description":5459,"org":5460,"tags":5461,"stars":20,"repoUrl":21,"updatedAt":5472},"accelerated-computing-cudf","accelerate data processing with cuDF","Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV\u002FParquet I\u002FO, nullable semantics, and multi-GPU DataFrame workloads.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5462,5465,5468,5469],{"name":5463,"slug":5464,"type":15},"Data Analysis","data-analysis",{"name":5466,"slug":5467,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":5470,"slug":5471,"type":15},"Performance","performance","2026-07-14T05:28:43.176466",{"slug":5474,"name":5474,"fn":5475,"description":5476,"org":5477,"tags":5478,"stars":20,"repoUrl":21,"updatedAt":5486},"aiq-deploy","deploy and manage NVIDIA AI-Q infrastructure","Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5479,5482,5485],{"name":5480,"slug":5481,"type":15},"Deployment","deployment",{"name":5483,"slug":5484,"type":15},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":5488,"name":5488,"fn":5489,"description":5490,"org":5491,"tags":5492,"stars":20,"repoUrl":21,"updatedAt":5500},"aiq-research","conduct deep research with AI-Q","Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5493,5496,5497],{"name":5494,"slug":5495,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":5498,"slug":5499,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":5502,"name":5502,"fn":5503,"description":5504,"org":5505,"tags":5506,"stars":20,"repoUrl":21,"updatedAt":5512},"amc-run-sample-calibration","run AMC sample dataset calibration","Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5507,5508,5509],{"name":5463,"slug":5464,"type":15},{"name":9,"slug":8,"type":15},{"name":5510,"slug":5511,"type":15},"Testing","testing","2026-07-17T05:29:03.913266",{"slug":5514,"name":5514,"fn":5515,"description":5516,"org":5517,"tags":5518,"stars":20,"repoUrl":21,"updatedAt":5529},"amc-run-video-calibration","calibrate video datasets with AutoMagicCalib","Calibrate a new dataset from pre-recorded video files via the AutoMagicCalib REST API. Use when user has local MP4s and says 'calibrate my videos', 'run AMC on these videos', or similar. For RTSP\u002Flive streams, use amc-run-rtsp-calibration instead.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5519,5522,5525,5526],{"name":5520,"slug":5521,"type":15},"Automation","automation",{"name":5523,"slug":5524,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":5527,"slug":5528,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":5531,"name":5531,"fn":5532,"description":5533,"org":5534,"tags":5535,"stars":20,"repoUrl":21,"updatedAt":5544},"amc-setup-calibration-stack","deploy AutoMagicCalib microservice with Docker","Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5536,5537,5540,5541],{"name":5480,"slug":5481,"type":15},{"name":5538,"slug":5539,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":5542,"slug":5543,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":5546,"name":5546,"fn":5547,"description":5548,"org":5549,"tags":5550,"stars":20,"repoUrl":21,"updatedAt":5558},"cudaq-guide","develop quantum applications with CUDA-Q","CUDA-Q onboarding guide for installation, test programs, GPU simulation, QPU hardware, and quantum applications.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5551,5552,5555],{"name":9,"slug":8,"type":15},{"name":5553,"slug":5554,"type":15},"Quantum Computing","quantum-computing",{"name":5556,"slug":5557,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305,{"items":5561,"total":5709},[5562,5580,5595,5606,5618,5632,5645,5657,5668,5677,5691,5700],{"slug":5563,"name":5563,"fn":5564,"description":5565,"org":5566,"tags":5567,"stars":5577,"repoUrl":5578,"updatedAt":5579},"nemoclaw-user-guide","retrieve NemoClaw documentation and configuration","Guides human users' AI agents to the NemoClaw docs MCP server and canonical Fern documentation in Markdown form. Use when users ask how to install, configure, operate, troubleshoot, secure, or learn NemoClaw with an AI coding assistant. Trigger keywords - nemoclaw docs, use nemoclaw with ai agent, nemoclaw mcp docs, nemoclaw install help, nemoclaw quickstart, nemoclaw markdown docs, llms.txt, agent skills.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5568,5571,5574],{"name":5569,"slug":5570,"type":15},"Documentation","documentation",{"name":5572,"slug":5573,"type":15},"MCP","mcp",{"name":5575,"slug":5576,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":5581,"name":5581,"fn":5582,"description":5583,"org":5584,"tags":5585,"stars":5592,"repoUrl":5593,"updatedAt":5594},"mcore-build-and-dependency","manage Megatron-LM development environments","Container-based dev environment setup and dependency management for Megatron-LM. Covers acquiring and launching the CI container, uv package management, and updating uv.lock.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5586,5589,5590],{"name":5587,"slug":5588,"type":15},"Containers","containers",{"name":5480,"slug":5481,"type":15},{"name":5591,"slug":618,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":5596,"name":5596,"fn":5597,"description":5598,"org":5599,"tags":5600,"stars":5592,"repoUrl":5593,"updatedAt":5605},"mcore-bump-base-image","update NVIDIA PyTorch base images","Bump the NVIDIA PyTorch base image (`nvcr.io\u002Fnvidia\u002Fpytorch:YY.MM-py3`) used by Megatron-LM CI. Covers the two pin sites (GitHub CI in `docker\u002F.ngc_version.dev` and GitLab CI in `.gitlab\u002Fstages\u002F01.build.yml`), the post-bump CI loop (re-run functional tests, refresh golden values, mark broken tests), and the gotchas that bit PRs",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5601,5604],{"name":5602,"slug":5603,"type":15},"CI\u002FCD","ci-cd",{"name":5480,"slug":5481,"type":15},"2026-07-14T05:25:59.97109",{"slug":5607,"name":5607,"fn":5608,"description":5609,"org":5610,"tags":5611,"stars":5592,"repoUrl":5593,"updatedAt":5617},"mcore-cicd","manage CI\u002FCD pipelines for Megatron-LM","CI\u002FCD reference for Megatron-LM. Covers CI pipeline structure, PR scope labels, triggering internal GitLab CI (which force-pushes the current branch to a pull-request\u002FBRANCH ref — always dry-run and verify the destination first; never run against shared or protected branches), and CI failure investigation.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5612,5613,5614],{"name":5602,"slug":5603,"type":15},{"name":5480,"slug":5481,"type":15},{"name":5615,"slug":5616,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":5619,"name":5619,"fn":5620,"description":5621,"org":5622,"tags":5623,"stars":5592,"repoUrl":5593,"updatedAt":5631},"mcore-create-issue","investigate CI failures and create issues","Investigate a failing GitHub Actions run or job and create a GitHub issue for the failure.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5624,5627,5628],{"name":5625,"slug":5626,"type":15},"Debugging","debugging",{"name":5615,"slug":5616,"type":15},{"name":5629,"slug":5630,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":5633,"name":5633,"fn":5634,"description":5635,"org":5636,"tags":5637,"stars":5592,"repoUrl":5593,"updatedAt":5644},"mcore-linting-and-formatting","lint and format Megatron-LM code","Linting and formatting for Megatron-LM. Covers running autoformat.sh, tools (ruff, black, isort, pylint, mypy), and code style rules.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5638,5641],{"name":5639,"slug":5640,"type":15},"Best Practices","best-practices",{"name":5642,"slug":5643,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":5646,"name":5646,"fn":5647,"description":5648,"org":5649,"tags":5650,"stars":5592,"repoUrl":5593,"updatedAt":5656},"mcore-migrate-gpt-to-hybrid","migrate Megatron-LM models to HybridModel","Migration guide for moving Megatron Core GPTModel checkpoints, model providers, training commands, and layer mappings to HybridModel.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5651,5652,5655],{"name":17,"slug":18,"type":15},{"name":5653,"slug":5654,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":5658,"name":5658,"fn":5659,"description":5660,"org":5661,"tags":5662,"stars":5592,"repoUrl":5593,"updatedAt":5667},"mcore-onboard-gb200-1node-tests","onboard functional tests for GB200","Onboard 1-node GitHub MR functional tests for GB200 from existing mr-scoped 2-node tests.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5663,5666],{"name":5664,"slug":5665,"type":15},"QA","qa",{"name":5510,"slug":5511,"type":15},"2026-07-14T05:25:53.673039",{"slug":5669,"name":5669,"fn":5670,"description":5671,"org":5672,"tags":5673,"stars":5592,"repoUrl":5593,"updatedAt":5676},"mcore-run-on-slurm","launch distributed training jobs on SLURM","How to launch distributed Megatron-LM training jobs on a SLURM cluster. Covers a minimal sbatch skeleton, environment-variable setup for torch.distributed.run, CUDA_DEVICE_MAX_CONNECTIONS rules across hardware and parallelism modes, container conventions, monitoring, and per-rank failure diagnosis.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5674,5675],{"name":5480,"slug":5481,"type":15},{"name":5483,"slug":5484,"type":15},"2026-07-14T05:25:49.362534",{"slug":5678,"name":5678,"fn":5679,"description":5680,"org":5681,"tags":5682,"stars":5592,"repoUrl":5593,"updatedAt":5690},"mcore-split-pr","split pull requests to reduce review load","Split a PR into multiple PRs to reduce the number of required CODEOWNERS reviewer groups.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5683,5686,5687],{"name":5684,"slug":5685,"type":15},"Code Review","code-review",{"name":5615,"slug":5616,"type":15},{"name":5688,"slug":5689,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":5692,"name":5692,"fn":5693,"description":5694,"org":5695,"tags":5696,"stars":5592,"repoUrl":5593,"updatedAt":5699},"mcore-testing","run and manage Megatron-LM tests","Test system for Megatron-LM. Covers test layout, recipe YAML structure, adding and running unit and functional tests, golden values, marker filters, and CI parity.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5697,5698],{"name":5664,"slug":5665,"type":15},{"name":5510,"slug":5511,"type":15},"2026-07-14T05:25:54.928983",{"slug":5701,"name":5701,"fn":5702,"description":5703,"org":5704,"tags":5705,"stars":5592,"repoUrl":5593,"updatedAt":5708},"nightly-sync","manage nightly main-to-dev sync workflows","Domain knowledge for the nightly main-to-dev sync workflow. Covers merge strategy, CI architecture, failure investigation, and known issues.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[5706,5707],{"name":5520,"slug":5521,"type":15},{"name":5602,"slug":5603,"type":15},"2026-07-30T05:29:03.275638",496]