Fix dequant mixed for qwen3.5 quantized model made by vllm/llm-compressor (#4675)

* fix dequant_mixed

* add llm_config check

* remove dequantize_gemm

* add CompressedTensorFormat.dequant

* Potential fix for pull request finding

* Potential fix for pull request finding

Author: irexyc

lmdeploy/turbomind/linear.py

@@ -73,8 +73,8 @@ def _dequant_linear(linear: Linear, *, data_type) -> Linear:
     """Dequantize a quantized Linear to trivial.
 
     ``TrivialFormat.dequant`` is identity, so already-trivial inputs round-trip
-    safely.  ``AWQFormat.dequant`` and ``FP8Format.dequant`` do real work.
-    GPTQ / CompressedTensor / MXFP4 inherit the base-class
+    safely.  ``AWQFormat.dequant``, ``CompressedTensorFormat.dequant`` and
+    ``FP8Format.dequant`` do real work. GPTQ / MXFP4 inherit the base-class
     ``NotImplementedError`` — calling ``_dequant_linear`` on one of those is a
     broken-fusion-group configuration, and the raise names it at the call site.
     """

lmdeploy/turbomind/weight_format.py

@@ -75,9 +75,9 @@ def pack_u4_row(x: torch.Tensor) -> torch.Tensor:
 
 
 def _zeros_int4_symmetric(scales: Tensor) -> Tensor:
-    """Synthesize symmetric int4 zero-points (value = 8) matching *scales*
-    shape."""
-    return torch.full(scales.shape, 8, dtype=torch.uint8, device=scales.device)
+    """Synthesize normalized symmetric int4 zero-points (value = 8) matching
+    *scales* shape."""
+    return torch.full(scales.shape, 8, dtype=scales.dtype, device=scales.device)
 
 
 # ---------------------------------------------------------------------------
@@ -329,6 +329,22 @@ def pack(self, tensor: Tensor, kind: str) -> PackedTensor:
     def synthesize_zeros(self, scales: Tensor) -> Tensor:
         return _zeros_int4_symmetric(scales)
 
+    def dequant(self, tensors, data_type):
+        weight = tensors['weight']
+        scales = tensors['scales']
+        zeros = tensors['zeros']
+
+        out_size = weight.shape[-1]
+        zeros = zeros[..., :out_size]
+
+        scales = scales.repeat_interleave(self.block_in, dim=0)[:weight.shape[0]]
+        zeros = zeros.repeat_interleave(self.block_in, dim=0)[:weight.shape[0]]
+        w = (weight.to(scales.dtype) - zeros.to(scales.dtype)) * scales
+        result: dict[str, Tensor] = {'weight': w}
+        if 'bias' in tensors:
+            result['bias'] = tensors['bias']
+        return result
+
 
 class FP8Format(WeightFormat):
     name           = 'fp8'