[FIX_FOR_VLLM_CUSTOM=dcacdf9a8860a86401127d1c8f93ebf3cfbfd026] Fix MultiModelEngineClient, Qwen3.5 compilation, and EPLB refactoring#1436
Merged
iboiko-habana merged 10 commits intoMay 19, 2026
Conversation
pawel-olejniczak
requested review from
PatrykWo,
adobrzyn,
afierka-intel,
iboiko-habana,
jbyczkow,
kamil-kaczor,
ksmusz,
mgawarkiewicz-intel,
michalkuligowski and
xuechendi
as code owners
🚧 CI BlockedThe main CI workflow was not started for the following reason:
|
Contributor
There was a problem hiding this comment.
Pull request overview
This PR aims to restore compatibility with upstream vLLM changes and to avoid an HPU torch.compile failure for the Qwen3.5-35B-A3B evaluation path by forcing eager execution.
Changes:
- Implement
EngineClient.notify_kv_transfer_request_rejected()inMultiModelEngineClientby delegating to the underlying engine. - Update the Qwen3.5-35B-A3B full-test model card to set
enforce_eager: true(intended to bypass graph compilation).
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
vllm_gaudi/entrypoints/openai/multi_model_api_server.py |
Adds delegation method to satisfy upstream EngineClient abstract API and includes minor formatting-only adjustments. |
tests/full_tests/model_cards/qwen3.5-35b-a3b.yaml |
Adds enforce_eager: true to attempt to disable compile/graph capture for this model’s full-test run. |
pawel-olejniczak
force-pushed
the
fix/vllm-hourly-ci-fixes
branch
3 times, most recently
from
4a53ec6 to
81b2daa
Compare
🚧 CI BlockedThe main CI workflow was not started for the following reason:
|
pawel-olejniczak
changed the title
pawel-olejniczak
changed the title
pawel-olejniczak
force-pushed
the
fix/vllm-hourly-ci-fixes
branch
2 times, most recently
from
c355fb7 to
6d45ba9
Compare
…ltiModelEngineClient abstract method and Qwen3.5 compilation - Add notify_kv_transfer_request_rejected() delegation to MultiModelEngineClient (upstream PR #41269 added new abstract method to EngineClient) - Set enforce_eager=true for Qwen3.5-35B-A3B model card to work around aot_autograd view mutation assertion that fires during HPU graph compilation (upstream compilation changes between vLLM 8eb40113 and 9efdddca trigger incompatibility with HPU's monkey-patched attention) Signed-off-by: Paweł Olejniczak <pawelx.olejniczak@intel.com>
The enforce_eager parameter was only read from the ENFORCE_EAGER environment variable, ignoring the value set in model card YAML files. This caused the Qwen3.5-35B-A3B test to fail with a BackendCompilerFailed error on HPU because torch.compile was not disabled despite enforce_eager: true being set in the model card. Read enforce_eager from eval_config (model card) first, with env var as override — consistent with how trust_remote_code, dtype, and other model config fields are handled. Signed-off-by: Paweł Olejniczak <pawelx.olejniczak@intel.com>
pawel-olejniczak
force-pushed
the
fix/vllm-hourly-ci-fixes
branch
from
6d45ba9 to
4488cec
Compare
pawel-olejniczak
changed the title
pawel-olejniczak
changed the title
pawel-olejniczak
force-pushed
the
fix/vllm-hourly-ci-fixes
branch
from
4488cec to
35b43be
Compare
… EMPTY_EPLB_STATE import and enable_eplb parameter after upstream EPLB refactor Signed-off-by: Paweł Olejniczak <pawelx.olejniczak@intel.com>
pawel-olejniczak
force-pushed
the
fix/vllm-hourly-ci-fixes
branch
from
35b43be to
b8231a4
Compare
shepark
reviewed
May 14, 2026
Proper fix for Qwen3.5 compilation (mamba_type Enum comparison) is in PR vllm-project#1449. The enforce_eager workaround causes performance degradation and is unnecessary once vllm-project#1449 merges. Signed-off-by: Pawel Olejniczak <pawelx.olejniczak@intel.com> Signed-off-by: Paweł Olejniczak <pawelx.olejniczak@intel.com>
🚧 CI BlockedThe main CI workflow was not started for the following reason:
|
iboiko-habana
approved these changes
May 19, 2026
✅ CI PassedAll checks passed successfully against the following vllm commit: |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fix upstream regressions affecting hourly CI:
notify_kv_transfer_request_rejectedabstract method (upstream PR [Bugfix][KV Transfer][NIXL] Notify P node on pre-admission rejection to free stranded KV blocks vllm#41269)test_common.pyto readenforce_eagerfrom model card config (with env var override), enabling per-model compilation controlEMPTY_EPLB_STATEimport andenable_eplbparameter frompatched_create_fused_moe_routerafter upstream MoE refactor (upstream PR [MoE Refactor] EPLB refactoring for FusedMoE vllm#41055)Note: The
enforce_eager: trueworkaround for Qwen3.5 compilation has been removed — the root cause (mamba_type str-vs-Enum comparison in hybrid cache allocation) is properly fixed by #1449, which should merge first.Verified on HPU: unit tests pass on Gaudi 3 (MoE, FP8, compressed tensors).