[Feat.][310P]: weightNZ feature with quant or unquant. (#6705)

NZ Format Support for Linear Layers: Implemented support for the NZ
(N-dimensional Z-order) format for linear layer weights on Ascend 310P,
enhancing performance for both quantized and unquantized layers.
Unquantized Linear Method for Ascend 310P: Introduced
AscendUnquantizedLinearMethod310 to specifically handle and apply NZ
format casting to unquantized linear layer weights during the loading
process.
MRotaryEmbedding Integration: Extended Rotary Embedding support by
adding AscendMRotaryEmbedding310 to provide an Ascend-specific
implementation for MRotaryEmbedding.
Quantization Method Updates: Updated the w8a8_static quantization method
to directly transpose weights and apply NZ format casting, ensuring
consistency with the new format.
- vLLM version: v0.15.0
- vLLM main:
9562912cea

---------

Signed-off-by: Tflowers-0129 <2906339855@qq.com>

vllm_ascend/_310p/quantization/methods/w8a8_static.py

@@ -21,6 +21,7 @@ import torch
 import torch_npu
 
 from vllm_ascend.quantization.methods.base import AscendLinearScheme
+from vllm_ascend.utils import ACL_FORMAT_FRACTAL_NZ
 
 from .registry import register_scheme
 
@@ -72,9 +73,15 @@ class AscendW8A8LinearMethod310(AscendLinearScheme):
 
         quant_bias = layer.quant_bias if tp_rank == 0 else None
 
+        # NOTE(310P):
+        # - Current torch_npu.npu_quant_matmul on Ascend 310P expects the weight layout in a transposed form
+        #   for correct/efficient execution, so we pass `layer.weight.T` here.
+        # - This is a temporary workaround. The planned replacement quant-matmul op will accept the
+        #   canonical (non-transposed) weight layout directly, so this explicit transpose will be removed
+        #   once that op is enabled on 310P.
         return torch_npu.npu_quant_matmul(
             x,
-            layer.weight,
+            layer.weight.data,
             layer.deq_scale,
             bias=quant_bias,
             output_dtype=layer.params_dtype,
@@ -82,6 +89,8 @@ class AscendW8A8LinearMethod310(AscendLinearScheme):
 
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
         expanding_factor = layer.weight.data.shape[1]
+
+        # ---- quant stage tensors ----
         layer.aclnn_input_scale = torch.nn.Parameter(
             layer.input_scale.data.repeat(expanding_factor),
             requires_grad=False,
@@ -95,7 +104,9 @@ class AscendW8A8LinearMethod310(AscendLinearScheme):
             requires_grad=False,
         ).to(layer.aclnn_input_scale.dtype)
 
-        layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()
+        # ---- matmul stage tensor ----
+        layer.weight.data = torch_npu.npu_format_cast(layer.weight.data, ACL_FORMAT_FRACTAL_NZ).transpose(0, 1)
 
+        # ---- dequant stage tensors ----
         layer.weight_scale.data = torch.flatten(layer.weight_scale.data)
         layer.weight_offset.data = torch.flatten(layer.weight_offset.data)

vllm_ascend/_310p/quantization/modelslim_config.py

@@ -104,9 +104,9 @@ class AscendModelSlimConfig310(AscendModelSlimConfig):
         if isinstance(layer, LinearBase):
             packed = getattr(self, "packed_modules_mapping", {})
             if self.is_layer_skipped_ascend(prefix, packed):
-                from vllm_ascend.ops.linear import AscendUnquantizedLinearMethod
+                from vllm_ascend._310p.ops.linear import AscendUnquantizedLinearMethod310
 
-                return AscendUnquantizedLinearMethod()
+                return AscendUnquantizedLinearMethod310()
 
             scheme = create_scheme_for_layer(
                 quant_description=self.quant_description,