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] 2026-01-20 02:35:08 +00:00 • committed by Jai Suphavadeeprasit
parent ee23761709
commit a4142a73bc
1 changed files with 14 additions and 14 deletions
Download patch file Download diff file Expand all files Collapse all files

28
example_trainer/README.md
View file

@ -691,13 +691,13 @@ The JSON file contains everything needed to reconstruct tensor references in ano
"model": "Qwen/Qwen2.5-3B-Instruct",
"tp_degree": 1,
"dp_shard_degree": 1,
"param_names": [
"model.embed_tokens.weight",
"model.layers.0.self_attn.qkv_proj.weight",
...
],
"param_mappings": {
"model.embed_tokens.weight": {
"vllm_name": "model.embed_tokens.weight",
@ -707,23 +707,23 @@ The JSON file contains everything needed to reconstruct tensor references in ano
},
...
},
"ipc_handles": {
"model.embed_tokens.weight": {
"device_index": 0,
"ipc_handle_b64": "AmPA0pN...",
"ipc_handle_b64": "AmPA0pN...",
"storage_size": 623902720,
"storage_offset": 0,
"ref_counter_handle_b64": "Y2JY...",
"ref_counter_offset": 0,
"event_handle_b64": "wRIs...",
"event_handle_b64": "wRIs...",
"event_sync_required": true,
"shape": [152064, 2048],
"dtype": "torch.bfloat16"
},
...
},
"shared_weights_enabled": true,
"single_copy_enabled": true,
"num_params": 255
@ -756,15 +756,15 @@ The JSON file contains everything needed to reconstruct tensor references in ano
for name, ipc_info in config["ipc_handles"].items():
# Decode IPC handle from base64
ipc_handle = base64.b64decode(ipc_info["ipc_handle_b64"])
# Reconstruct storage from IPC handle
storage = torch.UntypedStorage._new_shared_cuda(
device_index, ipc_handle, storage_size, ...
)
# Create tensor from shared storage
tensor = torch.tensor(storage).view(shape).to(dtype)
# Replace model parameter with shared tensor
model.get_parameter(name).data = tensor
```
@ -900,7 +900,7 @@ pkill -9 -u $USER -f "vllm|grpo|python|run-api"
## Feature Availability Matrix
### What's Available
### What's Available
| Feature | Status | Notes |
|---------|--------|-------|
@ -916,7 +916,7 @@ pkill -9 -u $USER -f "vllm|grpo|python|run-api"
| **Wandb Logging** | Working | Via `--use-wandb` flag |
| **Custom Environments** | Working | Extend `BaseEnv` class |
### What's NOT Available
### What's NOT Available
| Feature | Mode | Status | Reason / Workaround |
|---------|------|--------|---------------------|
@ -942,7 +942,7 @@ pkill -9 -u $USER -f "vllm|grpo|python|run-api"
| **LoRA** | Supported | Via vLLM | Multiple Trainers |
| **Legacy** | Supported | Via vLLM | Multiple Trainers |
> **Key Point**: The multi-GPU limitation is **ONLY for single-copy mode** due to CUDA IPC constraints.
> **Key Point**: The multi-GPU limitation is **ONLY for single-copy mode** due to CUDA IPC constraints.
> LoRA and Legacy modes work with standard vLLM which fully supports tensor parallelism.
#### Pipeline Parallel (PP)
@ -1040,7 +1040,7 @@ CUDA_VISIBLE_DEVICES=5 python -u example_trainer/grpo.py \
## Future Work
### High Priority
### High Priority
| Feature | Description |
|---------|-------------|
@ -1048,7 +1048,7 @@ CUDA_VISIBLE_DEVICES=5 python -u example_trainer/grpo.py \
| **Automatic Server Type Detection** | Auto-detect correct `server_type` for environments |
| **Checkpoint Resume** | Resume training from checkpoints seamlessly |
### Medium Priority
### Medium Priority
| Feature | Description | Difficulty |
|---------|-------------|------------|