fix(weight_utils): name weight_type companions after the last weight only

[KaigeGao1110]

Purpose

The GGUF weight iterators name the synthetic weight_type companion with name.replace("weight", "weight_type"), which rewrites every occurrence. A module whose own name contains weight therefore loses its name.

Qwen3.8-Flash-Next's hyper-connection projection input_mix_weight_down.weight got the companion input_mix_weight_type_down.weight_type. No module owns that name, so loading the GGUF failed with no module or parameter named hyper_connection_mixer.input_mix_weight_type_down.

Changes

• Add gguf_weight_type_name() in weight_utils.py. It replaces only the last weight (the parameter name) and raises if the name has none.
• Use it in both the LLM iterator (gguf_quant_weights_iterator_multi) and the diffusion iterator (weights_adapter/diffusion/base.py).

Names with a single weight are unchanged, so existing models map exactly as before.

Test plan

tests/test_weight_type_names.py:

• the helper keeps the module name for single-weight names, *_weight.weight names and fused w13_weight, and rejects names without weight;
• both real iterators, run over a written Q8_0 GGUF file, yield the companion name derived from the last weight only.

With the two call sites reverted to str.replace (the helper kept), the iterator tests fail:

FAILED tests/test_weight_type_names.py::test_iterator_names_companions_after_the_mapped_parameter
  'model.hyper_connection_mixer.input_mix_weight_type_down.weight_type' != 'model.hyper_connection_mixer.input_mix_weight_down.weight_type'
FAILED tests/test_weight_type_names.py::test_diffusion_iterator_names_companions_after_the_parameter
  'transformer_blocks.0.attn.gate_weight_type_proj.weight_type' != 'transformer_blocks.0.attn.gate_weight_proj.weight_type'
2 failed, 6 passed

With the fix, all 8 tests pass.

Test results

I ran the full suite on one RTX PRO 6000 Blackwell (SM120) with vLLM nightly 2a02f6ef (0.28.1rc1.dev628) and PyTorch 2.13.0+cu130, with Hugging Face offline. I ran it on this branch and on unmodified main (d4c1f0d), using the same _C_gguf build (csrc/ is untouched):

tree	failed	passed	skipped
main	81	1552	578
this PR	81	1560	578

The new tests pass (8 more passed). The 81 failures are the same test IDs on both trees, and none of them comes from this change:

• 68 × test_kernels.py::test_moe: triton.runtime.errors.OutOfResources: out of resource: shared memory, because the Triton block size exceeds this GPU's limit.
• 7 × test_gguf_generation.py::test_models and 5 × test_multimodal_gguf.py: the models cannot be downloaded offline (OSError, LocalEntryNotFoundError, and HFValidationError for repo:quant references).
• 1 × test_plugin.py::test_register_sets_engine_args_for_gguf_model: HFValidationError for the local path /tmp/model.gguf.

pre-commit run --all-files passes.

Duplicate check: I searched open and closed PRs for weight_type. Nothing touches companion naming.

🤖 Generated with Claude Code