Support llm-compressor symmetric quantized model inference in TurboMind

[43758726]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily receiving feedbacks. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Make symmetric awq/gptq model quantized by llm-compressor can be inferenced in turboMind.

Modification

lmdeploy/lmdeploy/turbomind/deploy/parameter.py: add module that need to init weight_zero_point if model quantized from llm-compressor is symmetric.

Use cases (Optional)

from lmdeploy import pipeline, TurbomindEngineConfig
engine_config = TurbomindEngineConfig()
pipe = pipeline("{awq/gptq model path quantzied by llm-compressor}", 
                backend_config=engine_config)
response = pipe(["Hi, pls intro yourself", "Shanghai is"])
print(response)

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
4. The documentation has been modified accordingly, like docstring or example tutorials.

[Copilot]

Pull request overview

Adds support for symmetric AWQ/GPTQ models quantized by llm-compressor when running inference with TurboMind by ensuring a zeros (zero-point) tensor exists for compressed weights.

Changes:

• Add a fallback path to generate weight_zero_point tensors when missing (intended for symmetric quantized compressed-tensors models).
• Adjust Parameter.take() to return the matched key list, enabling CompressedWeight to detect whether weight_zero_point exists.
• Update get_params() wiring to pass matched keys into CompressedWeight.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

generate_zero_point() relies on g('weight_shape'), but weight_shape is not a supported kind in the input policies (e.g. process_compressed_tensor) and many readers (e.g. LlamaReader._attn) use dict.get, so this can return None or a tensor of shape values. In either case, row, col = weight_shape will produce tensors/None and torch.full((row, col // 128), ...) will raise at runtime. Prefer deriving the zeros tensor shapes from g('weight_scale') (or g('weight_packed')) which are guaranteed present for CompressedWeight, and allocate on the same device as the other tensors.

Suggested change

	import torch
	weight_shapes = g('weight_shape')
	result = []
	for weight_shape in weight_shapes:
	row, col = weight_shape
	tensor = torch.full((row, col // 128), 8, dtype=torch.uint8)
	weight_scales = g('weight_scale')
	# Normalize to a tuple of tensors for uniform handling
	if isinstance(weight_scales, torch.Tensor):
	weight_scales = (weight_scales, )
	result = []
	for scale in weight_scales:
	# Match the shape and device of the corresponding scale tensor
	tensor = torch.full(scale.shape, 8, dtype=torch.uint8, device=scale.device)

[Copilot]

This import of module torch is redundant, as it was previously imported on line 5.
This import of module lmdeploy.pytorch.check_env.torch is redundant, as it was previously imported on line 5.

Suggested change

import torch