[Ilya-a-sergeyev-ger]

What

should_convert_module in src/transformers/quantizers/quantizers_utils.py matches each modules_to_not_convert entry against the module's full name with an unanchored re.match(f"{key}", full_name).

Because re.match anchors at the start but not the end, a literal pattern such as ...mlp.gate also matches ...mlp.gate_proj. When gate_proj is not supposed to be converted, the result is inverted: gate_proj gets skipped when gate was the intended target (or vice versa), so the quantizer leaves it as a plain nn.Linear while its weight is FP8 — a BF16 × FP8 matmul then dies at forward with mat1 BFloat16 != mat2 Float8_e4m3fn.

Repro

Real checkpoint: Qwen/Qwen3.8-27B-FP8 (quant_method "fp8"), whose config lists model.language_model.layers.N.mlp.gate among 882 modules_to_not_convert entries. The MoE router mlp.gate must not be quantized, but mlp.gate_proj must be.

• re.match("model.language_model.layers.0.mlp.gate", "model.language_model.layers.0.mlp.gate_proj.weight") → True (bug: skips gate_proj)
• After anchoring: re.match(f"{key}$", full_name) → False, gate_proj converts correctly.

Fix

Anchor the regex to exact match, preserving the existing layer.* wildcard semantics:

re.match(f"{key}\\.", full_name) or re.match(f"{key}$", full_name) or full_name.endswith(key)

A regression test is added to tests/quantization/mxfp4/test_mxfp4.py::test_should_convert_module asserting that a literal pattern does not skip a longer sibling that only shares its prefix.

Why this matters

It is not a niche cosmetic issue: any finegrained-FP8 MoE checkpoint whose config names a router submodule that is a prefix of another module (e.g. mlp.gate vs mlp.gate_proj) either fails to load or silently mis-quantizes. The current logic has no way to express "exactly this module" for such names.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: mxfp4

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 31897425256:2
Result: success | Jobs: 16 | Tests: 181,622 | Failures: 0 | Duration: 17h 48m

[IlyasMoutawwakil]

hey, thanks for the report and the repro! the gate/gate_proj case is a real bug, but anchoring with $ stops matching regex entries that end in their own boundary group: ^model\.audio_tower(?:\.|$) no longer covers model.audio_tower.layers.1.proj, and the same unanchored prefix also lets model.layers.1 skip layers 10-19. #49771 makes a match from the start end at a module boundary (a consumed ., or right before a . or the end of the name), which fixes both and keeps those regex entries; it covers your regression case. would you be ok closing this in favor of it? happy to add you as co-author there.