fix(pack): forward quant_result to make_quant in pack_model

[Anai-Guo]

Problem

make_quant declares quant_result as its third required positional parameter:

https://github.com/ModelCloud/GPTQModel/blob/main/gptqmodel/utils/model.py#L398-L409

def make_quant(
    module,
    qcfg: QuantizeConfig,
    quant_result: Dict[str, Dict[str, Any]],
    backend: BACKEND,
    lm_head_name: str,
    ...

but the call inside pack_model never passes it:

quant_linear_cls = make_quant(
    model,
    qcfg=qcfg,
    backend=backend,          # <-- quant_result never supplied
    lm_head_name=lm_head_name,
    pack=True,
    device=DEVICE.CPU,
)

So every call to pack_model fails immediately with:

TypeError: make_quant() missing 1 required positional argument: 'quant_result'

Why this is an omission rather than intended

The value is already in scope and is used correctly everywhere else in the very same function — pack_model takes quant_result as its own second required parameter, and:

• line 1167 — modules = {n: modules[n] for n in quant_result} (three lines above the broken call)
• line 1180 — if name in quant_result when building qModules
• line 1213 — pack_module(..., quant_result=quant_result, ...)

The two other make_quant call sites in the codebase both pass it:

call site	passes quant_result?
gptqmodel/models/loader.py:1418	✅ quant_result=modules
gptqmodel/models/writer.py:1171	✅ quant_result=modules
gptqmodel/utils/model.py:1168 (pack_model)	❌ — this PR

Fix

One line — forward the parameter that is already in scope:

     quant_linear_cls = make_quant(
         model,
         qcfg=qcfg,
+        quant_result=quant_result,
         backend=backend,

Verification

I do not have a GPU box here, so rather than hand-copying a repro (which can silently diverge from the source) I replayed the actual argument binding straight out of the AST: make_quant's parameter list and each Call node's positional/keyword shape are read from the real files, a signature-identical stub is generated, and the binding is replayed.

make_quant signature reconstructed from source:
    (module, qcfg, quant_result, backend, lm_head_name, pack=None, device=None,
     from_quantized=None, dtype=None, is_sharded=None)

UNPATCHED -- callsite gptqmodel/utils/model.py:1168 (inside pack_model)
  passes 1 positional + keywords ['qcfg', 'backend', 'lm_head_name', 'pack', 'device']
  pack_model -> make_quant           TypeError: make_quant() missing 1 required
                                     positional argument: 'quant_result'

SIBLING CALLSITES (the baseline this one deviates from)
  loader.py:1418                     BINDS OK
  writer.py:1171                     BINDS OK

PATCHED -- add quant_result=quant_result to the pack_model callsite
  pack_model -> make_quant (fixed)   BINDS OK

The unpatched run reproduces the exact TypeError from the report above, and the patched run binds cleanly.

Note on the base commit

My fork could not be synced to main (the GitHub API refuses merge-upstream for this token because the sync would touch .github/workflows/code_quality.yml, which needs the workflow OAuth scope). The branch is therefore based on my fork's own tip. I checked the drift explicitly: the make_quant call block in pack_model is byte-identical between that commit and current main, and the upstream changes to pack_model since then (the qModules comprehension and the GPTQMODEL_MAX_PACKERS handling) are all below the lines this patch touches. The resulting diff is exactly +1/-0 against main. Happy to rebase if you'd prefer.

---

🤖 Generated with Claude Code