[Quark][MTP] Only drop the draft quant config when the MTP experts are excluded

[jiaqiang-dot-liu]

Problem

_mtp_quant_config drops the Quark quant config for the whole MTP draft module if
any mtp.* entry appears in the checkpoint's exclude list:

if any(isinstance(layer, str) and layer.startswith("mtp.") for layer in exclude_layers):
    return None

That is right for checkpoints that ship the entire MTP module in bf16, which is the
case the current comment describes (sgl-project#23113).

It is wrong for checkpoints that exclude only the dense sub-modules. The
Qwen3.8-2.4T-A95B Quark MXFP4 export excludes mtp.fc,
mtp.layers.0.self_attn.*, mtp.layers.0.mlp.gate,
mtp.layers.0.mlp.shared_expert* — while still shipping MXFP4-packed
mtp.layers.*.mlp.experts.* tensors.

For those, returning None makes the fused-MoE loader allocate bf16
[N, hidden] and then fail copying the packed uint8 [N, hidden // 2]
checkpoint tensor.

Change

Only drop the quant config when the MTP experts are excluded too:

mtp_excluded = [l for l in exclude_layers if isinstance(l, str) and l.startswith("mtp.")]
if mtp_excluded and any("mlp.experts" in l for l in mtp_excluded):
    return None

Checkpoints that exclude the whole MTP module still match (their exclude list
contains the expert entries), so the sgl-project#23113 case is unchanged.

Verification

Not tested. No GPU or Python interpreter here, and I do not have either
checkpoint to check the exclude lists against.

The shape mismatch described above — bf16 [N, hidden] versus packed uint8
[N, hidden // 2] — is a 2x on the last dimension and matches the failure mode,
but I am taking the description of the Qwen3.8 export's exclude list on trust.

Before merging: dump quantization_config.exclude for both a whole-module-excluded
checkpoint and a dense-only-excluded one, and confirm the predicate classifies each
correctly.

Note on the substring match

"mlp.experts" in layer is a substring test, so it also matches a hypothetical
mtp.layers.0.mlp.experts_something. A stricter path-segment match would be
tighter; I kept the patch's form since it mirrors how the surrounding code already
matches prefixes.

Related

Five other agent sessions produced variants of this same fix
(_quark_mtp_per_tensor_exclude, _mtp_quark_scoped_exclude_gate,
_qwen3_5_mtp_quark_partial_exclude_gate x2,
_qwen3_5_mtp_quark_routed_expert_quant_gate), using regex segment matching or a
should_ignore_layer probe instead. This is the one that applies cleanly and makes
the smallest change; the others are recorded but not submitted.

---

Provenance: originally authored by an automated kernel-optimization agent
(hyperloom session 20260816T070000Z-ish, model Qwen3.8-2.4T-A95B-Quark-MXFP4);
applies cleanly to current main.

---

CI States

Latest PR Test (Base): ❌ Run #34923950761
Latest PR Test (Extra): ❌ Run #34923950432
Latest PR Test (AMD ROCm 10): ❌ Run #34923950926

[jiaqiang-dot-liu]

Superseded upstream: sgl-project#39064 (0154f72) landed the same predicate (drop the MTP quant config only when mtp.*.mlp.experts are excluded), plus packed_modules_mapping and a unit test.