NousResearch/atropos
1
0
Fork 0
mirror of https://github.com/NousResearch/atropos.git synced 2026-04-19 12:57:58 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

[pre-commit.ci] auto fixes from pre-commit.com hooks

Browse source 
for more information, see https://pre-commit.ci
This commit is contained in:
pre-commit-ci[bot] 2026-02-20 04:58:43 +00:00
parent ccdd5a1ca6
commit 60fb6cae11
11 changed files with 221 additions and 136 deletions
Download patch file Download diff file Expand all files Collapse all files

20
example_trainer/config.py
View file

@ -136,15 +136,17 @@ class TrainingConfig(BaseModel):
wandb_group: Optional[str] = Field(None, description="Wandb group name")
# === Training Mode Configuration ===
weight_bridge_mode: Literal["shared_vllm", "lora_only", "lora_restart", "none"] = Field(
"none",
description=(
"How to synchronize weights with inference server. "
"'shared_vllm': attach to vLLM's shared memory tensors and update in-place. "
"'lora_only': keep base model frozen, train/swap LoRA adapters via HTTP (slow, needs --enforce-eager). "
"'lora_restart': LoRA training with vLLM restarts (fast, CUDA graphs enabled). "
"'none': legacy mode, restart vLLM with new checkpoint files."
),
weight_bridge_mode: Literal["shared_vllm", "lora_only", "lora_restart", "none"] = (
Field(
"none",
description=(
"How to synchronize weights with inference server. "
"'shared_vllm': attach to vLLM's shared memory tensors and update in-place. "
"'lora_only': keep base model frozen, train/swap LoRA adapters via HTTP (slow, needs --enforce-eager). "
"'lora_restart': LoRA training with vLLM restarts (fast, CUDA graphs enabled). "
"'none': legacy mode, restart vLLM with new checkpoint files."
),
)
)
# === Distributed Training Configuration ===