[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>

docs/platforms/ascend_npu_deepseek_example.md

@@ -22,7 +22,6 @@ export SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1
 #npu acceleration operator
 export SGLANG_NPU_USE_MLAPO=1
 export SGLANG_USE_FIA_NZ=1
-export ENABLE_MOE_NZ=1
 
 python3 -m sglang.launch_server \
     --model-path ${MODEL_PATH} \
@@ -71,7 +70,6 @@ export HCCL_BUFFSIZE=1536
 #npu acceleration operator
 export SGLANG_NPU_USE_MLAPO=1
 export SGLANG_USE_FIA_NZ=1
-export ENABLE_MOE_NZ=1
 export TASK_QUEUE_ENABLE=2
 
 python -m sglang.launch_server \

docs/platforms/ascend_npu_qwen3_examples.md

@@ -62,7 +62,6 @@ export HCCL_BUFFSIZE=1536
 export HCCL_OP_EXPANSION_MODE=AIV
 export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=32
 export SGLANG_DEEPEP_BF16_DISPATCH=1
-export ENABLE_ASCEND_MOE_NZ=1
 
 python -m sglang.launch_server \
    --device npu \
@@ -84,7 +83,6 @@ export STREAMS_PER_DEVICE=32
 export HCCL_BUFFSIZE=1536
 export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=32
 export SGLANG_DEEPEP_BF16_DISPATCH=1
-export ENABLE_ASCEND_MOE_NZ=1
 
 python -m sglang.launch_server \
    --model-path Qwen/Qwen3-235B-A22B-Instruct-2507 \

python/sglang/srt/hardware_backend/npu/quantization/fused_moe_method_npu.py

@@ -150,42 +150,27 @@ class _NPUFusedMoEMethodBase(FusedMoEMethodBase):
 
 class NPUW8A8Int8DynamicMoEMethod(_NPUFusedMoEMethodBase):
 
-    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 process_weights_after_loading(self, layer: torch.nn.Module) -> None:
-        weight_data = self._release_weight_cache(layer.w13_weight.data)
+        layer.w13_weight.data = npu_format_cast(layer.w13_weight.data.transpose(1, 2))
-        layer.w13_weight = torch.nn.Parameter(weight_data, requires_grad=False)
+        layer.w2_weight.data = npu_format_cast(layer.w2_weight.data.transpose(1, 2))
-
-        weight_data = self._release_weight_cache(layer.w2_weight.data)
-        layer.w2_weight = torch.nn.Parameter(weight_data, requires_grad=False)
-
         layer.w13_weight_scale = torch.nn.Parameter(
-            layer.w13_weight_scale.data.squeeze(-1).contiguous().to(torch.float32),
+            layer.w13_weight_scale.data.squeeze(-1), requires_grad=False
-            requires_grad=False,
         )
         layer.w2_weight_scale = torch.nn.Parameter(
-            layer.w2_weight_scale.data.squeeze(-1).contiguous(), requires_grad=False
+            layer.w2_weight_scale.data.squeeze(-1), requires_grad=False
         )
         # Compressed-tensors format doesn't have this field
         if hasattr(layer, "w13_weight_offset"):
             layer.w13_weight_offset = torch.nn.Parameter(
-                layer.w13_weight_offset.data.squeeze(-1).contiguous(),
+                layer.w13_weight_offset.data.squeeze(-1),
                 requires_grad=False,
             )
         if hasattr(layer, "w2_weight_offset"):
             layer.w2_weight_offset = torch.nn.Parameter(
-                layer.w2_weight_offset.data.squeeze(-1).contiguous(),
+                layer.w2_weight_offset.data.squeeze(-1),
                 requires_grad=False,
             )
 
-        layer.w13_weight.data = npu_format_cast(layer.w13_weight.data)
-        layer.w2_weight.data = npu_format_cast(layer.w2_weight.data)
-
     def apply(
         self,
         layer,

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)
+        cpu_w13 = layer.w13_weight.transpose(1, 2).cpu()
-        torch_npu.npu_format_cast_(w13, 2)
-        cpu_w13 = w13.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,

python/sglang/srt/models/qwen3_moe.py

@@ -71,6 +71,7 @@ from sglang.srt.models.utils import (
 )
 from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
+    LazyValue,
     add_prefix,
     is_cuda,
     is_flashinfer_available,
@@ -1119,14 +1120,16 @@ class Qwen3MoeForCausalLM(nn.Module):
                     else:
                         logger.warning(f"Parameter {name} not found in params_dict")
 
-        # TODO mimic deepseek
-        # Lazy initialization of expert weights cache to avoid slowing down load_weights
         if not hasattr(self, "routed_experts_weights_of_layer"):
-            self.routed_experts_weights_of_layer = {
+            self.routed_experts_weights_of_layer = LazyValue(
-                layer_id: self.model.layers[layer_id].mlp.get_moe_weights()
+                lambda: {
-                for layer_id in range(self.start_layer, self.end_layer)
+                    layer_id: self.model.layers[layer_id].mlp.get_moe_weights()
-                if isinstance(self.model.layers[layer_id].mlp, Qwen3MoeSparseMoeBlock)
+                    for layer_id in range(self.start_layer, self.end_layer)
-            }
+                    if isinstance(
+                        self.model.layers[layer_id].mlp, Qwen3MoeSparseMoeBlock
+                    )
+                }
+            )
 
     @classmethod
     def get_model_config_for_expert_location(cls, config):

test/registered/ascend/test_ascend_memory_consumption.py

@@ -0,0 +1,76 @@
+"""
+Usage:
+python3 -m unittest test_ascend_memory_consumption.TestMemoryConsumptionAscend.test_memory_consumption
+"""
+
+import os
+import unittest
+
+import torch
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_npu_ci
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+
+if "ASCEND_RT_VISIBLE_DEVICES" not in os.environ:
+    os.environ["ASCEND_RT_VISIBLE_DEVICES"] = "0,1"
+DEFAULT_PORT_FOR_SRT_TEST_RUNNER = (
+    8000 + int(os.environ.get("ASCEND_RT_VISIBLE_DEVICES", "0")[0]) * 100
+)
+DEFAULT_URL_FOR_TEST = f"http://127.0.0.1:{DEFAULT_PORT_FOR_SRT_TEST_RUNNER + 1000}"
+
+
+class TestMemoryConsumptionAscend(CustomTestCase):
+
+    def test_memory_consumption(self):
+
+        model = "nytopop/Qwen3-30B-A3B.w8a8"
+        base_url = DEFAULT_URL_FOR_TEST
+
+        ### Calculate initial used memory
+        free_npu_memory, total_npu_memory = torch.npu.mem_get_info()
+        initial_used_memory = total_npu_memory - free_npu_memory
+
+        process = popen_launch_server(
+            model,
+            base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--trust-remote-code",
+                "--device",
+                "npu",
+                "--attention-backend",
+                "ascend",
+                "--tp-size",
+                "2",
+                "--mem-fraction-static",
+                "0.8",
+                "--cuda-graph-bs",
+                "1",
+                "--max-total-tokens",
+                "1024",
+                "--disable-radix-cache",
+                "--disable-cuda-graph",
+            ],
+        )
+
+        ### Calculate initial used memory
+        free_npu_memory, total_npu_memory = torch.npu.mem_get_info()
+        used_memory_after_server_starting = (
+            total_npu_memory - free_npu_memory - initial_used_memory
+        ) / (1 << 30)
+        self.assertLessEqual(float(used_memory_after_server_starting), 16.00)
+
+        # Clean up everything
+        kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()