[NPU] NZ for non-quantized MOE, Qwen3 MOE double memory consumption fix (#15904)

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

python/sglang/srt/layers/moe/ep_moe/layer.py

@@ -7,6 +7,7 @@ import torch
 
 from sglang.srt.compilation.piecewise_context_manager import is_in_piecewise_cuda_graph
 from sglang.srt.environ import envs
+from sglang.srt.hardware_backend.npu.utils import npu_format_cast
 from sglang.srt.layers import deep_gemm_wrapper
 from sglang.srt.layers.moe import (
     get_deepep_mode,
@@ -472,13 +473,6 @@ class NpuFuseEPMoE(DeepEPMoE):
             gmm2_weight_scale=self.w2_weight_scale,
         ).hidden_state
 
-    def release_weight_cache(self, weight: torch.Tensor):
-        # .contiguous() introduces additional memory overhead and needs to be released using resize_(0)
-        origin_weight = weight.data.transpose(1, 2)
-        new_weight = origin_weight.contiguous()
-        origin_weight.untyped_storage().resize_(0)
-        return new_weight
-
     def permute_w13_weight_scale(self, w: torch.Tensor, tile_n: int):
         if tile_n % 2 != 0:
             raise ValueError(f"tile_n must be even, got {tile_n}")
@@ -520,14 +514,12 @@ class NpuFuseEPMoE(DeepEPMoE):
         return weight.view(*original_shape[:dim], -1, *original_shape[dim + 1 :])
 
     def _process_weights_after_loading(self, layer: torch.nn.Module) -> None:
-        w13 = self.release_weight_cache(layer.w13_weight)
-        torch_npu.npu_format_cast_(w13, 2)
-        cpu_w13 = w13.cpu()
+        cpu_w13 = layer.w13_weight.transpose(1, 2).cpu()
         w13 = self.reshape_w13_weight(cpu_w13, -1).npu()
-        torch_npu.npu_format_cast_(w13, 29)
+        w13 = npu_format_cast(w13)
         layer.w13_weight = torch.nn.Parameter(w13, requires_grad=False)
 
-        w2 = torch_npu.npu_format_cast(layer.w2_weight.data, 29)
+        w2 = npu_format_cast(layer.w2_weight)
         layer.w2_weight = torch.nn.Parameter(w2, requires_grad=False)
 
         w13_scale = layer.w13_weight_scale.data.squeeze(-1).contiguous()

python/sglang/srt/layers/quantization/unquant.py

@@ -25,6 +25,7 @@ from sglang.srt.utils import (
     get_bool_env_var,
     is_cpu,
     is_hip,
+    is_npu,
     next_power_of_2,
     set_weight_attrs,
     use_intel_amx_backend,
@@ -40,6 +41,7 @@ if TYPE_CHECKING:
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_hip = is_hip()
 _is_cpu = is_cpu()
+_is_npu = is_npu()
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
 if _use_aiter:
@@ -47,6 +49,9 @@ if _use_aiter:
     from aiter.fused_moe import fused_moe
     from aiter.ops.shuffle import shuffle_weight
 
+if _is_npu:
+    from sglang.srt.hardware_backend.npu.utils import npu_format_cast
+
 try:
     from flashinfer.fused_moe import cutlass_fused_moe as flashinfer_cutlass_fused_moe
 except ImportError:
@@ -296,6 +301,14 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
                 layer.num_local_experts, *new_shape_w2
             )
 
+        if _is_npu:
+            for weight_name in ["w13_weight", "w2_weight"]:
+                weight = getattr(layer, weight_name)
+                weight.data = weight.data.transpose(1, 2)
+                weight.data = npu_format_cast(
+                    weight.data,
+                )
+
         return
 
     def create_moe_runner(
@@ -494,14 +507,11 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
         expert_tokens = expert_tokens.to(torch.int64)
         w13_bias = [layer.w13_weight_bias] if self.with_bias else None
         w2_bias = [layer.w2_weight_bias] if self.with_bias else None
-        if layer.w13_weight.shape[-1] == layer.hidden_size:
-            w13 = layer.w13_weight.transpose(1, 2)
-            w2 = layer.w2_weight.transpose(1, 2)
 
         # gmm1: gate_up_proj
         hidden_states = torch_npu.npu_grouped_matmul(
             x=[hidden_states],
-            weight=[w13],
+            weight=[layer.w13_weight],
             bias=w13_bias,
             split_item=2,
             group_list_type=0,
@@ -525,7 +535,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
         # gmm2: down_proj
         hidden_states = torch_npu.npu_grouped_matmul(
             x=[hidden_states],
-            weight=[w2],
+            weight=[layer.w2_weight],
             bias=w2_bias,
             split_item=2,
             group_list_type=0,