[Quantization] Introduce layer re-quantization API for linear layers (MXFP8 linear showcase)

[fxmarty-amd]

Disclosure

AI assistance was used. The changes were reviewed and tested manually.

Purpose

Draft PR adapted from vllm-project#48427 integrating with vllm-project#51392 internal API.

This change lets --quantization-config.linear fp8_per_channel convert selected serialized ModelOpt MXFP8 linears to FP8 PTPC at load time, namely this relaxes online quantization to be available for pre-quantized layers if a dequantize_weights method is available in the base quant method.

The original checkpoint quantization config remains responsible for unselected linears and MoE.

The composition is handled by QuantizationConfig.get_effective_quant_method, so model implementations continue to receive their original checkpoint quantization config type.

Quantization methods expose an optional dequantize_weight hook, that:

• raises NotImplementedError by default, or
• dequantizes the quant method weights if implemented.

This draft implements it only for ModelOptMxFp8LinearMethod.

Such dequantize_weight methods can be shared across model producers (quark, modelopt, compressed-tensors, etc.). No model producer-specific code should be required.

The online linear base accepts a requantization source, while FP8 PTPC is the only target that consumes it in this draft.

TODO:

• Validate rejection on non-linear layers

Test Plan

vllm serve mmangkad/Qwen3-4B-Instruct-2507-MXFP8 --quantization-config.linear "fp8_per_channel"

or

vllm serve MiniMaxAI/MiniMax-M3-MXFP8 \
  --tensor-parallel-size 4 \
  --attention-backend TRITON_ATTN \
  --moe-backend aiter \
  --linear-backend auto \
  --kv-cache-dtype fp8 \
  --no-enable-prefix-caching \
  --language-model-only \
  --quantization-config.linear fp8_per_channel

pytest tests/quantization/test_online_shared_expert_fusion.py -s -vvvvv

Test Result

Unit test passing, end-to-end eval to be added.

[fxmarty-amd]

See vllm-project#55684