fix: support Qwen3.6 DFlash and cleanup SGLang patching#131
Conversation
Signed-off-by: Doğaç Eldenk <dogacel@gmail.com>
ee3ce9d to
0a7b68b
Compare
0a7b68b to
b3b6bee
Compare
Signed-off-by: Doğaç Eldenk <dogacel@gmail.com>
b3b6bee to
04ed996
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0a7b68b75e
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| embedding_key: model.language_model.embed_tokens.weight | ||
| lm_head_key: lm_head.weight | ||
| norm_key: model.language_model.norm.weight | ||
| draft_model_config: torchspec/config/dflash_draft_config_qwen36_35b.json |
There was a problem hiding this comment.
Add the missing Qwen3.6 DFlash draft config
When this new Qwen3.6 config is used, train_entry._get_draft_model_config() passes this value directly to AutoDraftModelConfig.from_file(), which opens the path with no fallback. I searched the repo for dflash_draft_config_qwen36_35b.json and it is not present, so launching configs/sglang_qwen36_35b_dflash.yaml fails before training with FileNotFoundError instead of enabling the advertised DFlash setup.
Useful? React with 👍 / 👎.
|
|
||
| SGLANG_COMMIT=$(grep "^ARG SGLANG_COMMIT=" "$SGLANG_DIR/Dockerfile" | cut -d= -f2) | ||
| # shellcheck source=tools/sglang_lib.sh | ||
| source "$SCRIPT_DIR/sglang_lib.sh" |
There was a problem hiding this comment.
Honor the path passed to apply_sglang_patch.sh
After sourcing sglang_lib.sh, SGLANG_PATH is always derived from SGLANG_FOLDER_NAME (defaulting to ${PROJECT_ROOT}/_sglang), and the positional <path-to-sglang-repo> argument is never assigned anymore. In the documented invocation ./tools/apply_sglang_patch.sh /tmp/sglang, the script will still cd, git reset --hard, and git clean -fd the default _sglang checkout, ignoring the requested target and potentially modifying the wrong repository.
Useful? React with 👍 / 👎.
Overview
There is no need to store previous patches, SGLang 0.5.14 seems to be stable now. I also made some cleanup on the patch generation script and added a reference on how to use that in README.