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
2025-07-09 16:53:45 +00:00
parent f980cda286
commit 3e13ea033e
5 changed files with 282 additions and 246 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
search.json
View File

@@ -400,7 +400,7 @@
"href": "docs/multi-gpu.html#sec-deepspeed",
"title": "Multi-GPU",
"section": "2 DeepSpeed",
"text": "2 DeepSpeed\nDeepSpeed is the recommended approach for multi-GPU training due to its stability and performance. It provides various optimization levels through ZeRO stages.\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.",
"text": "2 DeepSpeed\nDeepSpeed is the recommended approach for multi-GPU training due to its stability and performance. It provides various optimization levels through ZeRO stages.\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.\n\n\n\n\n\n\n\n\nTip\n\n\n\nUsing ZeRO Stage 3 with Single-GPU training\nZeRO Stage 3 can be used for training on a single GPU by manually setting the environment variables:\nWORLD_SIZE=1 LOCAL_RANK=0 MASTER_ADDR=0.0.0.0 MASTER_PORT=29500",
"crumbs": [
"Deployments",
"Multi-GPU"
@@ -2791,14 +2791,14 @@
"href": "docs/api/monkeypatch.gradient_checkpointing.offload_cpu.html",
"title": "monkeypatch.gradient_checkpointing.offload_cpu",
"section": "",
"text": "monkeypatch.gradient_checkpointing.offload_cpu\nCPU offloaded checkpointing\n\n\n\n\n\nName\nDescription\n\n\n\n\nCPU_Offloaded_Gradient_Checkpointer\nSaves VRAM by smartly offloading to RAM.\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_cpu.CPU_Offloaded_Gradient_Checkpointer(\n)\nSaves VRAM by smartly offloading to RAM.\nTiny hit to performance, since we mask the movement via non blocking calls."
"text": "monkeypatch.gradient_checkpointing.offload_cpu\nCPU offloaded checkpointing\n\n\n\n\n\nName\nDescription\n\n\n\n\nCPU_Offloaded_Gradient_Checkpointer\nSaves VRAM by smartly offloading to RAM.\n\n\nCheckpointFunctionWithCPUOffload\nThis is a torch/utils/checkpoint.py CheckpointFunction monkey patch that offloads the first tensor to cpu during forward and back to cuda during backward. This allows significant memory savings when using a very long seqlen. e.g. for llama 8b at 100k it’s 24GB saved per gpu: ((100_000*4096)*2*32/2**30)\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_cpu.CPU_Offloaded_Gradient_Checkpointer(\n)\nSaves VRAM by smartly offloading to RAM.\nTiny hit to performance, since we mask the movement via non blocking calls.\n\n\n\nmonkeypatch.gradient_checkpointing.offload_cpu.CheckpointFunctionWithCPUOffload(\n)\nThis is a torch/utils/checkpoint.py CheckpointFunction monkey patch that offloads the first tensor to cpu during forward and back to cuda during backward. This allows significant memory savings when using a very long seqlen. e.g. for llama 8b at 100k it’s 24GB saved per gpu: ((100_000*4096)*2*32/2**30)\nIn the case of a very long seqlen 100k+ the copying to/from cpu overhead is not big, because dense quadratic attention compute will dominate."
},
{
"objectID": "docs/api/monkeypatch.gradient_checkpointing.offload_cpu.html#classes",
"href": "docs/api/monkeypatch.gradient_checkpointing.offload_cpu.html#classes",
"title": "monkeypatch.gradient_checkpointing.offload_cpu",
"section": "",
"text": "Name\nDescription\n\n\n\n\nCPU_Offloaded_Gradient_Checkpointer\nSaves VRAM by smartly offloading to RAM.\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_cpu.CPU_Offloaded_Gradient_Checkpointer(\n)\nSaves VRAM by smartly offloading to RAM.\nTiny hit to performance, since we mask the movement via non blocking calls."
"text": "Name\nDescription\n\n\n\n\nCPU_Offloaded_Gradient_Checkpointer\nSaves VRAM by smartly offloading to RAM.\n\n\nCheckpointFunctionWithCPUOffload\nThis is a torch/utils/checkpoint.py CheckpointFunction monkey patch that offloads the first tensor to cpu during forward and back to cuda during backward. This allows significant memory savings when using a very long seqlen. e.g. for llama 8b at 100k it’s 24GB saved per gpu: ((100_000*4096)*2*32/2**30)\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_cpu.CPU_Offloaded_Gradient_Checkpointer(\n)\nSaves VRAM by smartly offloading to RAM.\nTiny hit to performance, since we mask the movement via non blocking calls.\n\n\n\nmonkeypatch.gradient_checkpointing.offload_cpu.CheckpointFunctionWithCPUOffload(\n)\nThis is a torch/utils/checkpoint.py CheckpointFunction monkey patch that offloads the first tensor to cpu during forward and back to cuda during backward. This allows significant memory savings when using a very long seqlen. e.g. for llama 8b at 100k it’s 24GB saved per gpu: ((100_000*4096)*2*32/2**30)\nIn the case of a very long seqlen 100k+ the copying to/from cpu overhead is not big, because dense quadratic attention compute will dominate."
},
{
"objectID": "docs/api/core.trainers.mamba.html",