Skip to content

fix(cicd): auto-select free port and fix qwen3 FSDP2 attention_type compat#151

Open
kcz358 wants to merge 2 commits intomainfrom
fix/cicd-free-port-and-qwen3-attention-type
Open

fix(cicd): auto-select free port and fix qwen3 FSDP2 attention_type compat#151
kcz358 wants to merge 2 commits intomainfrom
fix/cicd-free-port-and-qwen3-attention-type

Conversation

@kcz358
Copy link
Copy Markdown
Collaborator

@kcz358 kcz358 commented Apr 6, 2026

Summary

  • Auto-select free port for CICD tests: Added find_free_port() in test/train/utils.py so torchrun no longer hardcodes port 8000, avoiding conflicts on shared machines.
  • Fix qwen3 FSDP2 attention_type compatibility: FSDP2-wrapped Qwen3DecoderLayer loses the attention_type attribute. Now uses config.layer_types[i] (new transformers 5.x API) with fallback to decoder_layer.attention_type for older versions.
  • Update uv.lock: Synced lock file to current environment (transformers 5.5.0, etc.).

Verified

All CICD tests pass with 2 GPUs:

  • qwen3_vl: 2/2 passed
  • qwen3: 2/2 passed
  • qwen2_5_vl: 3/3 passed

kcz358 added 2 commits April 5, 2026 20:13
…ompat

- Add find_free_port() in test/train/utils.py so CICD tests no longer
  hardcode port 8000, avoiding conflicts on shared machines.
- Fix AttributeError in qwen3_ops.py where FSDP2-wrapped decoder layers
  lose the attention_type attribute. Use config.layer_types[i] (new
  transformers API) with fallback to decoder_layer.attention_type.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant