feat: add Hygon DCU INT8 hipBLASLt GEMM path

[starrkk]

Summary

• add an int8-vllm-hygon-dcu MM weight backend using Hygon DCU hipBLASLt W8A8 channelwise GEMM
• support auto-quantized BF16 weights and optional selective BF16 fallback
• expose helpers for shared activation quantization across repeated projections

Why

This enables faster INT8 GEMM execution on Hygon DCU while keeping the path gated by the quantization scheme and clear dependency checks.

Validation

• branch rebuilt on latest ModelTC/LightX2V:main (89dfa833)
• git diff --check passed for the PR branch
• validated as part of the HunyuanVideo1.5 I2V 8-card benchmark path on Hygon DCU

[gemini-code-assist]

Code Review

This pull request introduces a selective BF16 fallback mechanism and integrates hipblaslt_w8a8_channelwise_gemm for Hygon DCU in the quantization pipeline, along with helper functions for managing weights, biases, and environment flags. It also updates module loading logic to support custom load functions. The review feedback highlights opportunities to optimize performance and robustness, specifically by caching the casted weight tensor in the BF16 fallback path to avoid redundant casting, removing redundant .contiguous() calls on already contiguous weights during GEMM execution, and safely checking for the existence of module.weight to prevent potential AttributeErrors when loading auto-quantized biases.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Casting a large weight tensor on every single forward pass is extremely inefficient and will cause significant GPU memory churn and latency overhead. We should cache the casted weight back to self.weight so that subsequent forward passes can reuse it directly.

Suggested change

	def _apply_bf16(self, input_tensor):
	weight = self.weight
	if weight.dtype != input_tensor.dtype:
	weight = weight.to(input_tensor.dtype)
	bias = _bias_or_none(self, input_tensor.dtype)
	return F.linear(input_tensor, weight, bias)
	def _apply_bf16(self, input_tensor):
	if self.weight.dtype != input_tensor.dtype:
	self.weight = self.weight.to(input_tensor.dtype)
	bias = _bias_or_none(self, input_tensor.dtype)
	return F.linear(input_tensor, self.weight, bias)

[gemini-code-assist]

To prevent a potential AttributeError if module.weight is None or not yet initialized, we should safely check for its existence and fallback to the bias tensor's own device.

Suggested change

	module.bias = bias.to(module.weight.device)
	device = module.weight.device if getattr(module, "weight", None) is not None else bias.device
	module.bias = bias.to(device)

[gemini-code-assist]

Since self.weight and self.weight_scale are already made contiguous during the load() phase (via _make_weight_contiguous), calling .contiguous() on them during every forward pass is redundant and adds unnecessary overhead.

Suggested change

	_, output_tensor = hipblaslt_gemm(
	a=input_tensor_quant.contiguous(),
	b=self.weight.contiguous(),
	scale_a=input_tensor_scale.contiguous(),
	scale_b=self.weight_scale.contiguous(),
	_, output_tensor = hipblaslt_gemm(
	a=input_tensor_quant.contiguous(),
	b=self.weight,
	scale_a=input_tensor_scale.contiguous(),
	scale_b=self.weight_scale,