fix(vllm): allow BF16 rollout with inherited quantization exclusions

[seonjinn]

Summary

BF16 rollout can inherit quantization_ignore_patterns from an MXFP8 recipe. The scope logger currently raises KeyError: 'quantization_config' before model initialization because BF16 has no generated quantization config.

Skip this diagnostic when its configuration is absent. Existing MXFP8 output and quantization-config validation are unchanged. No changes to weight loading, refit, or kernels.

Tests

• Four missing/null-config regression cases fail before the fix.
• GB200, vLLM 0.25.1: 256 generation/refit tests pass with this fix on the integration branch, including all 14 HF-override tests.
• Fresh 20-step BF16 rollout reruns are submitted; E2E results are pending.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[seonjinn]

/ok to test c49e043

[seonjinn]

Updated in d49af6c: explicit BF16 rollout with inherited quantization_ignore_patterns now emits a warning when no quantization configuration is active, then continues normally. Mixed MXFP8 rollout with BF16 first/last layers retains its effective-exclusion log and does not receive this warning. Added coverage for bf16/bfloat16 aliases, missing/None configuration, and mixed MXFP8. GB200 red/green regression validation has been submitted; the new warning tests are not yet claimed passing.

[seonjinn]

GB200 validation completed (job 7096575, exit 0). With the new warning assertions and the old implementation, 8 regression cases failed as expected. With the warning implementation, the full selected suite passed: 267 tests. This includes BF16 aliases with absent/None quantization config and mixed MXFP8 exclusions without a BF16 warning. Validation used integration source 1802f7c with the same helper/test changes as d49af6c; this is not a claim of every end-to-end configuration passing.