【feat】switch for fusion ops gmmswigluquant (#5992)

### What this PR does / why we need it?

Set a additional config parameter to control whether the gmmswigluequant
fuseion operator is enabled; it is enabled by True. / When enabled with
a small number of GPUs, the gmmswigluquant fused operator can cause some
performance degradation.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
2c24bc6996

#### Perf

test model: GLM 4.6(w8a8)
- single A3 node(ep16, tp16),  async-scheduling, mtp, FULL_DECODE_ONLY
- bs=1, input_lens=32000, ouput_lens=1024

Without this PR: TPOT 32.22.ms
With this PR: TPOT 30.23ms

---------

Signed-off-by: zjks98 <zhangjiakang4@huawei.com>
Co-authored-by: zjks98 <zhangjiakang4@huawei.com>

tests/ut/test_ascend_config.py

@@ -43,6 +43,9 @@ class TestAscendConfig(TestBase):
         ascend_compilation_config = ascend_config.ascend_compilation_config
         self.assertTrue(ascend_compilation_config.fuse_norm_quant)
 
+        ascend_fusion_config = ascend_config.ascend_fusion_config
+        self.assertTrue(ascend_fusion_config.fusion_ops_gmmswigluquant)
+
     @_clean_up_ascend_config
     def test_init_ascend_config_with_additional_config(self):
         test_vllm_config = VllmConfig()
@@ -50,6 +53,9 @@ class TestAscendConfig(TestBase):
             "ascend_compilation_config": {
                 "fuse_norm_quant": False,
             },
+            "ascend_fusion_config": {
+                "fusion_ops_gmmswigluquant": False,
+            },
             "multistream_overlap_shared_expert": True,
             "eplb_config": {
                 "num_redundant_experts": 2
@@ -66,6 +72,9 @@ class TestAscendConfig(TestBase):
         self.assertFalse(ascend_compilation_config.fuse_norm_quant)
         self.assertFalse(ascend_config.enable_kv_nz)
 
+        ascend_fusion_config = ascend_config.ascend_fusion_config
+        self.assertFalse(ascend_fusion_config.fusion_ops_gmmswigluquant)
+
     @_clean_up_ascend_config
     def test_init_ascend_config_enable_npugraph_ex(self):
         test_vllm_config = VllmConfig()

vllm_ascend/ascend_config.py

@@ -37,6 +37,9 @@ class AscendConfig:
         ascend_compilation_config = additional_config.get("ascend_compilation_config", {})
         self.ascend_compilation_config = AscendCompilationConfig(**ascend_compilation_config)
 
+        ascend_fusion_config = additional_config.get("ascend_fusion_config", {})
+        self.ascend_fusion_config = AscendFusionConfig(**ascend_fusion_config)
+
         finegrained_tp_config = additional_config.get("finegrained_tp_config", {})
         self.finegrained_tp_config = FinegrainedTPConfig(finegrained_tp_config, vllm_config)
 
@@ -190,6 +193,24 @@ class AscendCompilationConfig:
         self.fuse_allreduce_rms = fuse_allreduce_rms
 
 
+class AscendFusionConfig:
+    """
+    Configuration for controlling whether to use a fused operator gmmswigluquant.
+    """
+
+    def __init__(self, fusion_ops_gmmswigluquant: bool = True, **kwargs):
+        """
+        Initialize the configuration.
+
+        Args:
+            fusion_ops_gmmswigluquant (bool): Whether to use a fused operator gmmswigluquant.
+                When set to True, the system will use a fused operator gmmswigluquant.
+                Default: True
+            **kwargs: Additional optional parameters for forward compatibility and configuration extension.
+        """
+        self.fusion_ops_gmmswigluquant = fusion_ops_gmmswigluquant
+
+
 class XliteGraphConfig:
     """
     Configuration Object for xlite_graph_config from additional_config

vllm_ascend/ops/fused_moe/moe_comm_method.py

@@ -48,6 +48,12 @@ def setup_moe_comm_method(moe_config):
     _MoECommMethods[MoECommType.FUSED_MC2] = FusedMC2CommImpl(moe_config)
 
 
+def set_gmmswigluquant_method():
+    from vllm_ascend.ascend_config import get_ascend_config
+    ascend_config = get_ascend_config()
+    return ascend_config.ascend_fusion_config.fusion_ops_gmmswigluquant
+
+
 @dataclass
 class FusedExpertsResult:
     routed_out: torch.Tensor
@@ -69,6 +75,7 @@ class MoECommMethod(ABC):
 
         self.token_dispatcher = self._get_token_dispatcher()
         self.prepare_finalize = self._get_prepare_finalize()
+        self.use_fusion_ops = set_gmmswigluquant_method()
 
     def prepare(
         self,
@@ -159,7 +166,7 @@ class MoECommMethod(ABC):
             w2_offset=w2_offset,
             topk_scales=dispatch_results.topk_scales,
             with_quant=use_int8_w8a8 or use_int4_w4a8 or use_int4_w4a16,
-            fusion=use_int8_w8a8,
+            fusion=use_int8_w8a8 and self.use_fusion_ops,
             need_trans=need_trans,
             dynamic_eplb=dynamic_eplb)
 

vllm_ascend/platform.py

@@ -206,6 +206,13 @@ class NPUPlatform(Platform):
 
         elif model_config and hasattr(model_config.hf_text_config, "index_topk"):
             vllm_config.cache_config.cache_dtype = str(model_config.dtype).replace("torch.", "")
+
+        ascend_fusion_config = ascend_config.ascend_fusion_config
+        if ascend_fusion_config:
+            vllm_config.additional_config.setdefault("ascend_fusion_config", {}).update(
+                vars(ascend_fusion_config) if not isinstance(ascend_fusion_config, dict) else ascend_fusion_config
+            )
+
         if model_config is None:
             logger.warning("Model config is missing. This may indicate that we are running a test case")
             enforce_eager = False