tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Built site for gh-pages

Browse Source
This commit is contained in:
Quarto GHA Workflow Runner
2025-11-24 03:27:39 +00:00
parent 24cd17113a
commit cd7fdaeeb6
7 changed files with 436 additions and 420 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
search.json
View File

@@ -501,8 +501,8 @@
"objectID": "docs/multi-gpu.html#sec-overview",
"href": "docs/multi-gpu.html#sec-overview",
"title": "Multi-GPU",
"section": "1 Overview",
"text": "1 Overview\nAxolotl supports several methods for multi-GPU training:\n\nDeepSpeed (recommended)\nFSDP (Fully Sharded Data Parallel)\nSequence parallelism\nFSDP + QLoRA",
"section": "Overview",
"text": "Overview\nWhen training on multiple GPUs, Axolotl supports 3 sharding/parallelism strategies. Additionally, you can layer specific optimization features on top of that strategy.\nYou generally cannot combine these strategies; they are mutually exclusive.\n\nDeepSpeed: Powerful optimization library, supports ZeRO stages 1-3.\nFSDP (Fully Sharded Data Parallel): PyTorch’s native sharding implementation (Recommended).\nDDP (Distributed Data Parallel): PyTorch’s native parallelism implementation (Default if neither of the above are selected).\n\nThese features can often be combined with the strategies above:\n\nSequence Parallelism: Splits long sequences across GPUs (Compatible with DDP, DeepSpeed, and FSDP).\nFSDP + QLoRA: Combines 4-bit quantization with FSDP (Specific to FSDP).",
"crumbs": [
"Deployments",
"Multi-GPU"
@@ -512,8 +512,8 @@
"objectID": "docs/multi-gpu.html#sec-deepspeed",
"href": "docs/multi-gpu.html#sec-deepspeed",
"title": "Multi-GPU",
"section": "2 DeepSpeed",
"text": "2 DeepSpeed\n\n2.1 Configuration\nAdd to your YAML config:\ndeepspeed: deepspeed_configs/zero1.json\n\n\n2.2 Usage\n# Fetch deepspeed configs (if not already present)\naxolotl fetch deepspeed_configs\n\n# Passing arg via config\naxolotl train config.yml\n\n# Passing arg via cli\naxolotl train config.yml --deepspeed deepspeed_configs/zero1.json\n\n\n2.3 ZeRO Stages\nWe provide default configurations for:\n\nZeRO Stage 1 (zero1.json)\nZeRO Stage 1 with torch compile (zero1_torch_compile.json)\nZeRO Stage 2 (zero2.json)\nZeRO Stage 3 (zero3.json)\nZeRO Stage 3 with bf16 (zero3_bf16.json)\nZeRO Stage 3 with bf16 and CPU offload params(zero3_bf16_cpuoffload_params.json)\nZeRO Stage 3 with bf16 and CPU offload params and optimizer (zero3_bf16_cpuoffload_all.json)\n\n\n\n\n\n\n\nTip\n\n\n\nChoose the configuration that offloads the least amount to memory while still being able to fit on VRAM for best performance.\nStart from Stage 1 -> Stage 2 -> Stage 3.",
"section": "DeepSpeed",
"text": "DeepSpeed\n\nConfiguration\nAdd to your YAML config:\ndeepspeed: deepspeed_configs/zero1.json\n\n\nUsage\n# Fetch deepspeed configs (if not already present)\naxolotl fetch deepspeed_configs\n\n# Passing arg via config\naxolotl train config.yml\n\n# Passing arg via cli\naxolotl train config.yml --deepspeed deepspeed_configs/zero1.json\n\n\nZeRO Stages\nWe provide default configurations for:\n\nZeRO Stage 1 (zero1.json)\nZeRO Stage 1 with torch compile (zero1_torch_compile.json)\nZeRO Stage 2 (zero2.json)\nZeRO Stage 3 (zero3.json)\nZeRO Stage 3 with bf16 (zero3_bf16.json)\nZeRO Stage 3 with bf16 and CPU offload params(zero3_bf16_cpuoffload_params.json)\nZeRO Stage 3 with bf16 and CPU offload params and optimizer (zero3_bf16_cpuoffload_all.json)\n\n\n\n\n\n\n\nTip\n\n\n\nChoose the configuration that offloads the least amount to memory while still being able to fit on VRAM for best performance.\nStart from Stage 1 -> Stage 2 -> Stage 3.",
"crumbs": [
"Deployments",
"Multi-GPU"
@@ -523,8 +523,8 @@
"objectID": "docs/multi-gpu.html#sec-fsdp",
"href": "docs/multi-gpu.html#sec-fsdp",
"title": "Multi-GPU",
"section": "3 Fully Sharded Data Parallel (FSDP)",
"text": "3 Fully Sharded Data Parallel (FSDP)\n\n\n\n\n\n\nNote\n\n\n\nFSDP2 is recommended for new users. FSDP1 is deprecated and will be removed in an upcoming release of Axolotl.\n\n\n\n3.1 Migrating from FSDP1 to FSDP2\nTo migrate your config from FSDP1 to FSDP2, you must use the fsdp_version top-level config field to specify the FSDP version, and\nalso follow the config field mapping below to update field names.\n\n3.1.1 Config mapping\n\n\n\nFSDP1\nFSDP2\n\n\n\n\nfsdp_sharding_strategy\nreshard_after_forward\n\n\nfsdp_backward_prefetch_policy\nREMOVED\n\n\nfsdp_backward_prefetch\nREMOVED\n\n\nfsdp_forward_prefetch\nREMOVED\n\n\nfsdp_sync_module_states\nREMOVED\n\n\nfsdp_cpu_ram_efficient_loading\ncpu_ram_efficient_loading\n\n\nfsdp_state_dict_type\nstate_dict_type\n\n\nfsdp_use_orig_params\nREMOVED\n\n\nfsdp_activation_checkpointing\nactivation_checkpointing\n\n\n\nFor more details, please see the migration guide in the torchtitan repo. In Axolotl,\nif you were using the following FSDP1 config:\nfsdp_version: 1\nfsdp_config:\n fsdp_offload_params: false\n fsdp_cpu_ram_efficient_loading: true\n fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP\n fsdp_transformer_layer_cls_to_wrap: Qwen3DecoderLayer\n fsdp_state_dict_type: FULL_STATE_DICT\n fsdp_sharding_strategy: FULL_SHARD\nYou can migrate to the following FSDP2 config:\nfsdp_version: 2\nfsdp_config:\n offload_params: false\n cpu_ram_efficient_loading: true\n auto_wrap_policy: TRANSFORMER_BASED_WRAP\n transformer_layer_cls_to_wrap: Qwen3DecoderLayer\n state_dict_type: FULL_STATE_DICT\n reshard_after_forward: true\n\n\n\n3.2 FSDP1 (deprecated)\n\n\n\n\n\n\nNote\n\n\n\nUsing fsdp to configure FSDP is deprecated and will be removed in an upcoming release of Axolotl. Please use fsdp_config as above instead.\n\n\nfsdp:\n - full_shard\n - auto_wrap\nfsdp_config:\n fsdp_offload_params: true\n fsdp_state_dict_type: FULL_STATE_DICT\n fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer",
"section": "Fully Sharded Data Parallel (FSDP)",
"text": "Fully Sharded Data Parallel (FSDP)\nFSDP allows you to shard model parameters, gradients, and optimizer states across data parallel workers.\n\n\n\n\n\n\nNote\n\n\n\nFSDP2 is recommended for new users. FSDP1 is deprecated and will be removed in an upcoming release of Axolotl.\n\n\n\nFSDP + QLoRA\nFor combining FSDP with QLoRA, see our dedicated guide.\n\n\nMigrating from FSDP1 to FSDP2\nTo migrate your config from FSDP1 to FSDP2, you must use the fsdp_version top-level config field to specify the FSDP version, and\nalso follow the config field mapping below to update field names.\n\nConfig mapping\n\n\n\nFSDP1\nFSDP2\n\n\n\n\nfsdp_sharding_strategy\nreshard_after_forward\n\n\nfsdp_backward_prefetch_policy\nREMOVED\n\n\nfsdp_backward_prefetch\nREMOVED\n\n\nfsdp_forward_prefetch\nREMOVED\n\n\nfsdp_sync_module_states\nREMOVED\n\n\nfsdp_cpu_ram_efficient_loading\ncpu_ram_efficient_loading\n\n\nfsdp_state_dict_type\nstate_dict_type\n\n\nfsdp_use_orig_params\nREMOVED\n\n\nfsdp_activation_checkpointing\nactivation_checkpointing\n\n\n\nFor more details, please see the migration guide in the torchtitan repo. In Axolotl,\nif you were using the following FSDP1 config:\nfsdp_version: 1\nfsdp_config:\n fsdp_offload_params: false\n fsdp_cpu_ram_efficient_loading: true\n fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP\n fsdp_transformer_layer_cls_to_wrap: Qwen3DecoderLayer\n fsdp_state_dict_type: FULL_STATE_DICT\n fsdp_sharding_strategy: FULL_SHARD\nYou can migrate to the following FSDP2 config:\nfsdp_version: 2\nfsdp_config:\n offload_params: false\n cpu_ram_efficient_loading: true\n auto_wrap_policy: TRANSFORMER_BASED_WRAP\n transformer_layer_cls_to_wrap: Qwen3DecoderLayer\n state_dict_type: FULL_STATE_DICT\n reshard_after_forward: true\n\n\n\nFSDP1 (deprecated)\n\n\n\n\n\n\nNote\n\n\n\nUsing fsdp to configure FSDP is deprecated and will be removed in an upcoming release of Axolotl. Please use fsdp_config as above instead.\n\n\nfsdp:\n - full_shard\n - auto_wrap\nfsdp_config:\n fsdp_offload_params: true\n fsdp_state_dict_type: FULL_STATE_DICT\n fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer",
"crumbs": [
"Deployments",
"Multi-GPU"
@@ -534,8 +534,8 @@
"objectID": "docs/multi-gpu.html#sec-sequence-parallelism",
"href": "docs/multi-gpu.html#sec-sequence-parallelism",
"title": "Multi-GPU",
"section": "4 Sequence parallelism",
"text": "4 Sequence parallelism\nWe support sequence parallelism (SP) via the\nring-flash-attention project. This\nallows one to split up sequences across GPUs, which is useful in the event that a\nsingle sequence causes OOM errors during model training.\nSee our dedicated guide for more information.\n\n4.1 FSDP + QLoRA\nFor combining FSDP with QLoRA, see our dedicated guide.",
"section": "Sequence parallelism",
"text": "Sequence parallelism\nWe support sequence parallelism (SP) via the\nring-flash-attention project. This\nallows one to split up sequences across GPUs, which is useful in the event that a\nsingle sequence causes OOM errors during model training.\nSee our dedicated guide for more information.",
"crumbs": [
"Deployments",
"Multi-GPU"
@@ -545,8 +545,8 @@
"objectID": "docs/multi-gpu.html#sec-performance",
"href": "docs/multi-gpu.html#sec-performance",
"title": "Multi-GPU",
"section": "5 Performance Optimization",
"text": "5 Performance Optimization\n\n5.1 Liger Kernel Integration\nPlease see docs for more info.",
"section": "Performance Optimization",
"text": "Performance Optimization\n\nLiger Kernel Integration\nPlease see docs for more info.",
"crumbs": [
"Deployments",
"Multi-GPU"
@@ -556,8 +556,8 @@
"objectID": "docs/multi-gpu.html#sec-troubleshooting",
"href": "docs/multi-gpu.html#sec-troubleshooting",
"title": "Multi-GPU",
"section": "6 Troubleshooting",
"text": "6 Troubleshooting\n\n6.1 NCCL Issues\nFor NCCL-related problems, see our NCCL troubleshooting guide.\n\n\n6.2 Common Problems\n\nMemory IssuesTraining Instability\n\n\n\nReduce micro_batch_size\nReduce eval_batch_size\nAdjust gradient_accumulation_steps\nConsider using a higher ZeRO stage\n\n\n\n\nStart with DeepSpeed ZeRO-2\nMonitor loss values\nCheck learning rates\n\n\n\n\nFor more detailed troubleshooting, see our debugging guide.",
"section": "Troubleshooting",
"text": "Troubleshooting\n\nNCCL Issues\nFor NCCL-related problems, see our NCCL troubleshooting guide.\n\n\nCommon Problems\n\nMemory IssuesTraining Instability\n\n\n\nReduce micro_batch_size\nReduce eval_batch_size\nAdjust gradient_accumulation_steps\nConsider using a higher ZeRO stage\n\n\n\n\nStart with DeepSpeed ZeRO-2\nMonitor loss values\nCheck learning rates\n\n\n\n\nFor more detailed troubleshooting, see our debugging guide.",
"crumbs": [
"Deployments",
"Multi-GPU"
@@ -1910,7 +1910,7 @@
"href": "docs/custom_integrations.html#cut-cross-entropy",
"title": "Custom Integrations",
"section": "Cut Cross Entropy",
"text": "Cut Cross Entropy\nCut Cross Entropy (CCE) reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.\nSee https://github.com/apple/ml-cross-entropy\n\nRequirements\n\nPyTorch 2.4.0 or higher\n\n\n\nInstallation\nRun the following command to install cut_cross_entropy[transformers] if you don’t have it already.\n\nIf you are in dev environment\n\npython scripts/cutcrossentropy_install.py | sh\n\nIf you are installing from pip\n\npip3 uninstall -y cut-cross-entropy && pip3 install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@8a1a0ec\"\n\n\nUsage\nplugins:\n - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin\n\n\nSupported Models\n\napertus\narcee\ncohere\ncohere2\ndeepseek_v3\ngemma\ngemma2\ngemma3\ngemma3_text\ngemma3n\ngemma3n_text\nglm\nglm4\nglm4_moe\nglm4v\nglm4v_moe\ngpt_oss\ngranite\ngranitemoe\ngranitemoeshared\ngranitemoehybrid\nhunyuan_v1_dense\nhunyuan_v1_moe\nlfm2\nlfm2_moe\nlfm2_vl\nllama\nllama4\nllama4_text\nllava\nmistral\nmistral3\nmixtral\nmllama\nphi\nphi3\nphi4_multimodal\nqwen2\nqwen2_vl\nqwen2_moe\nqwen2_5_vl\nqwen3\nqwen3_moe\nqwen3_vl\nqwen3_vl_moe\nqwen3_next\nsmollm3\nseed_oss\nvoxtral\n\n\n\nCitation\n@article{wijmans2024cut,\n author = {Erik Wijmans and\n Brody Huval and\n Alexander Hertzberg and\n Vladlen Koltun and\n Philipp Kr\\\"ahenb\\\"uhl},\n title = {Cut Your Losses in Large-Vocabulary Language Models},\n journal = {arXiv},\n year = {2024},\n url = {https://arxiv.org/abs/2411.09009},\n}\nPlease see reference here",
"text": "Cut Cross Entropy\nCut Cross Entropy (CCE) reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.\nSee https://github.com/apple/ml-cross-entropy\n\nRequirements\n\nPyTorch 2.4.0 or higher\n\n\n\nInstallation\nRun the following command to install cut_cross_entropy[transformers] if you don’t have it already.\n\nIf you are in dev environment\n\npython scripts/cutcrossentropy_install.py | sh\n\nIf you are installing from pip\n\npip3 uninstall -y cut-cross-entropy && pip3 install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@5eff953\"\n\n\nUsage\nplugins:\n - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin\n\n\nSupported Models\n\napertus\narcee\ncohere\ncohere2\ndeepseek_v3\ngemma\ngemma2\ngemma3\ngemma3_text\ngemma3n\ngemma3n_text\nglm\nglm4\nglm4_moe\nglm4v\nglm4v_moe\ngpt_oss\ngranite\ngranitemoe\ngranitemoeshared\ngranitemoehybrid\nhunyuan_v1_dense\nhunyuan_v1_moe\nlfm2\nlfm2_moe\nlfm2_vl\nllama\nllama4\nllama4_text\nllava\nmistral\nmistral3\nmixtral\nmllama\nolmo\nolmo2\nolmo3\nphi\nphi3\nphi4_multimodal\nqwen2\nqwen2_vl\nqwen2_moe\nqwen2_5_vl\nqwen3\nqwen3_moe\nqwen3_vl\nqwen3_vl_moe\nqwen3_next\nsmollm3\nseed_oss\nvoxtral\n\n\n\nCitation\n@article{wijmans2024cut,\n author = {Erik Wijmans and\n Brody Huval and\n Alexander Hertzberg and\n Vladlen Koltun and\n Philipp Kr\\\"ahenb\\\"uhl},\n title = {Cut Your Losses in Large-Vocabulary Language Models},\n journal = {arXiv},\n year = {2024},\n url = {https://arxiv.org/abs/2411.09009},\n}\nPlease see reference here",
"crumbs": [
"Advanced Features",
"Custom Integrations"
@@ -2030,7 +2030,7 @@
"href": "index.html#latest-updates",
"title": "Axolotl",
"section": "🎉 Latest Updates",
"text": "🎉 Latest Updates\n\n2025/10: New model support has been added in Axolotl for: Qwen3 Next, Qwen2.5-vl, Qwen3-vl, Qwen3, Qwen3MoE, Granite 4, HunYuan, Magistral 2509, Apertus, and Seed-OSS.\n2025/09: Axolotl now has text diffusion training. Read more here.\n2025/08: QAT has been updated to include NVFP4 support. See PR.\n2025/07:\n\nND Parallelism support has been added into Axolotl. Compose Context Parallelism (CP), Tensor Parallelism (TP), and Fully Sharded Data Parallelism (FSDP) within a single node and across multiple nodes. Check out the blog post for more info.\nAxolotl adds more models: GPT-OSS, Gemma 3n, Liquid Foundation Model 2 (LFM2), and Arcee Foundation Models (AFM).\nFP8 finetuning with fp8 gather op is now possible in Axolotl via torchao. Get started here!\nVoxtral, Magistral 1.1, and Devstral with mistral-common tokenizer support has been integrated in Axolotl!\nTiledMLP support for single-GPU to multi-GPU training with DDP, DeepSpeed and FSDP support has been added to support Arctic Long Sequence Training. (ALST). See examples for using ALST with Axolotl!\n\n2025/05: Quantization Aware Training (QAT) support has been added to Axolotl. Explore the docs to learn more!\n\n\n\nExpand older updates\n\n\n2025/03: Axolotl has implemented Sequence Parallelism (SP) support. Read the blog and docs to learn how to scale your context length when fine-tuning.\n2025/06: Magistral with mistral-common tokenizer support has been added to Axolotl. See examples to start training your own Magistral models with Axolotl!\n2025/04: Llama 4 support has been added in Axolotl. See examples to start training your own Llama 4 models with Axolotl’s linearized version!\n2025/03: (Beta) Fine-tuning Multimodal models is now supported in Axolotl. Check out the docs to fine-tune your own!\n2025/02: Axolotl has added LoRA optimizations to reduce memory usage and improve training speed for LoRA and QLoRA in single GPU and multi-GPU training (DDP and DeepSpeed). Jump into the docs to give it a try.\n2025/02: Axolotl has added GRPO support. Dive into our blog and GRPO example and have some fun!\n2025/01: Axolotl has added Reward Modelling / Process Reward Modelling fine-tuning support. See docs.",
"text": "🎉 Latest Updates\n\n2025/11: Axolotl now includes support for Olmo3.\n2025/10: New model support has been added in Axolotl for: Qwen3 Next, Qwen2.5-vl, Qwen3-vl, Qwen3, Qwen3MoE, Granite 4, HunYuan, Magistral 2509, Apertus, and Seed-OSS.\n2025/09: Axolotl now has text diffusion training. Read more here.\n2025/08: QAT has been updated to include NVFP4 support. See PR.\n2025/07:\n\nND Parallelism support has been added into Axolotl. Compose Context Parallelism (CP), Tensor Parallelism (TP), and Fully Sharded Data Parallelism (FSDP) within a single node and across multiple nodes. Check out the blog post for more info.\nAxolotl adds more models: GPT-OSS, Gemma 3n, Liquid Foundation Model 2 (LFM2), and Arcee Foundation Models (AFM).\nFP8 finetuning with fp8 gather op is now possible in Axolotl via torchao. Get started here!\nVoxtral, Magistral 1.1, and Devstral with mistral-common tokenizer support has been integrated in Axolotl!\nTiledMLP support for single-GPU to multi-GPU training with DDP, DeepSpeed and FSDP support has been added to support Arctic Long Sequence Training. (ALST). See examples for using ALST with Axolotl!\n\n2025/05: Quantization Aware Training (QAT) support has been added to Axolotl. Explore the docs to learn more!\n\n\n\nExpand older updates\n\n\n2025/03: Axolotl has implemented Sequence Parallelism (SP) support. Read the blog and docs to learn how to scale your context length when fine-tuning.\n2025/06: Magistral with mistral-common tokenizer support has been added to Axolotl. See examples to start training your own Magistral models with Axolotl!\n2025/04: Llama 4 support has been added in Axolotl. See examples to start training your own Llama 4 models with Axolotl’s linearized version!\n2025/03: (Beta) Fine-tuning Multimodal models is now supported in Axolotl. Check out the docs to fine-tune your own!\n2025/02: Axolotl has added LoRA optimizations to reduce memory usage and improve training speed for LoRA and QLoRA in single GPU and multi-GPU training (DDP and DeepSpeed). Jump into the docs to give it a try.\n2025/02: Axolotl has added GRPO support. Dive into our blog and GRPO example and have some fun!\n2025/01: Axolotl has added Reward Modelling / Process Reward Modelling fine-tuning support. See docs.",
"crumbs": [
"Home"
]