Revert "CI: Use DeepSeek-R1-0528-mtp-mxfp4 models for deepseek fp4 tests"#206
Revert "CI: Use DeepSeek-R1-0528-mtp-mxfp4 models for deepseek fp4 tests"#206gyohuangxin wants to merge 1 commit intomainfrom
Conversation
There was a problem hiding this comment.
Pull request overview
Reverts #203 by restoring the previous DeepSeek FP4 model references in the ATOM CI workflow and adjusting how the Hugging Face model download step is executed.
Changes:
- Revert DeepSeek FP4/MTP
model_pathentries back toamd/DeepSeek-R1-0528-MXFP4. - Remove explicit
HF_TOKENinjection from thehf downloaddocker execstep.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| if [ -d "/models" ]; then | ||
| echo "/models directory found, downloading model to /models/${{ matrix.model_path }}" | ||
| if ! docker exec -e HF_TOKEN=${{ secrets.AMD_HF_TOKEN }} atom_test bash -lc "hf download ${{ matrix.model_path }} --local-dir /models/${{ matrix.model_path }}"; then | ||
| if ! docker exec atom_test bash -lc "hf download ${{ matrix.model_path }} --local-dir /models/${{ matrix.model_path }}"; then |
There was a problem hiding this comment.
The hf download invocation no longer passes HF_TOKEN (previously sourced from a secret) and the job does not set HF_TOKEN anywhere, so downloads will run unauthenticated unless the self-hosted runner environment provides HF_TOKEN. If any of these model repos are gated/private or hit rate limits, this step will fail; consider restoring token injection here or explicitly setting HF_TOKEN at the job/step level from an appropriate secret.
| - model_name: "DeepSeek-R1-0528-FP4" | ||
| model_path: "amd/DeepSeek-R1-0528-mtp-mxfp4" | ||
| model_path: "amd/DeepSeek-R1-0528-MXFP4" | ||
| extraArgs: "--kv_cache_dtype fp8 -tp 8" | ||
| env_vars: "" | ||
| runner: atom-mi355-8gpu.predownload | ||
| - model_name: "DeepSeek-R1-0528-FP4 MTP" | ||
| model_path: "amd/DeepSeek-R1-0528-mtp-mxfp4" | ||
| model_path: "amd/DeepSeek-R1-0528-MXFP4" | ||
| extraArgs: "--kv_cache_dtype fp8 -tp 8 --method mtp" |
There was a problem hiding this comment.
model_name labels these entries as "FP4", but the reverted model_path points at an "MXFP4" model repo. This mismatch makes CI matrix results harder to interpret/debug; consider aligning model_name with the actual repo name (or adding MXFP4 to the label) so logs/artifacts clearly indicate which model variant was tested.
Reverts #203