Improve torch compile for fused moe (#2327)

benchmark/kernels/fused_moe_triton/benchmark_torch_compile_fused_moe.py

@@ -6,6 +6,7 @@
 from transformers import AutoConfig
 
 from sglang.srt.layers.fused_moe_triton.fused_moe import fused_moe as fused_moe_triton
+from sglang.srt.model_executor.cuda_graph_runner import set_torch_compile_config
 
 
 def get_model_config(model_name: str, tp_size: int):
@@ -64,7 +65,7 @@ def fused_topk_native(
     return topk_weights, topk_ids
 
 
-@torch.compile
+@torch.compile(dynamic=False)
 def fused_moe_torch(
     x,
     w1,
@@ -88,7 +89,8 @@ def fused_moe_torch(
     w13_weights = w1[topk_ids]
     w1_weights, w3_weights = torch.chunk(w13_weights, 2, dim=2)
     w2_weights = w2[topk_ids]
-    x1 = F.gelu(torch.einsum("ti,taoi -> tao", x, w1_weights))
+    x1 = torch.einsum("ti,taoi -> tao", x, w1_weights)
+    x1 = F.silu(x1)
     x3 = torch.einsum("ti, taoi -> tao", x, w3_weights)
     expert_outs = torch.einsum("tao, taio -> tai", (x1 * x3), w2_weights)
     return torch.einsum("tai,ta -> ti", expert_outs, topk_weights.to(expert_outs.dtype))
@@ -174,6 +176,7 @@ def benchmark(batch_size, provider, model_config, use_fp8=False):
     print(f"benchmark {provider} with batch_size={batch_size}")
     torch.set_default_device("cuda")
     torch.cuda.manual_seed_all(0)
+    set_torch_compile_config()
 
     num_tokens = batch_size
     num_experts = model_config["num_experts"]

python/sglang/srt/layers/fused_moe_patch.py

@@ -105,20 +105,29 @@ def fused_moe_forward_native(
     num_expert_group: Optional[int] = None,
     custom_routing_function: Optional[Callable] = None,
 ) -> torch.Tensor:
-    assert custom_routing_function is None
-    topk_weights, topk_ids = select_experts_native(
-        hidden_states=x,
-        router_logits=router_logits,
-        use_grouped_topk=use_grouped_topk,
-        top_k=top_k,
-        renormalize=renormalize,
-        topk_group=topk_group,
-        num_expert_group=num_expert_group,
-    )
+
+    if use_grouped_topk:
+        assert num_expert_group is not None and topk_group is not None
+        topk_weights, topk_ids = grouped_topk(
+            x,
+            router_logits,
+            top_k,
+            renormalize,
+            num_expert_group,
+            topk_group,
+        )
+    elif custom_routing_function is None:
+        topk_weights, topk_ids = fused_topk_native(x, router_logits, top_k, renormalize)
+    else:
+        topk_weights, topk_ids = custom_routing_function(
+            x, router_logits, top_k, renormalize
+        )
+
     w13_weights = layer.w13_weight[topk_ids]
     w1_weights, w3_weights = torch.chunk(w13_weights, 2, dim=2)
     w2_weights = layer.w2_weight[topk_ids]
-    x1 = F.silu(torch.einsum("ti,taoi -> tao", x, w1_weights))
+    x1 = torch.einsum("ti,taoi -> tao", x, w1_weights)
+    x1 = F.silu(x1)
     x3 = torch.einsum("ti, taoi -> tao", x, w3_weights)
     expert_outs = torch.einsum("tao, taio -> tai", (x1 * x3), w2_weights)
     return torch.einsum("tai,ta -> ti", expert_outs, topk_weights.to(expert_outs.dtype))

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -36,7 +36,7 @@
     from sglang.srt.model_executor.model_runner import ModelRunner
 
 
-def _to_torch(model: torch.nn.Module, reverse: bool = False):
+def _to_torch(model: torch.nn.Module, reverse: bool, batch_size: int):
     for sub in model._modules.values():
         if isinstance(sub, CustomOp):
             if reverse:
@@ -45,38 +45,46 @@ def _to_torch(model: torch.nn.Module, reverse: bool = False):
             else:
                 # NOTE: Temporarily workaround MoE
                 if "FusedMoE" in sub.__class__.__name__:
-                    sub._forward_method = fused_moe_forward_native
+                    if batch_size == 1:
+                        # The performance of torch.compile on this layer is not always good when bs > 1,
+                        # so we decide to skip it for now.
+                        sub._forward_method = fused_moe_forward_native
                 else:
                     sub._forward_method = sub.forward_native
                 setattr(sub, "is_torch_compile", True)
         if isinstance(sub, torch.nn.Module):
-            _to_torch(sub, reverse)
+            _to_torch(sub, reverse, batch_size)
 
 
 @contextmanager
 def patch_model(
-    model: torch.nn.Module, enable_compile: bool, tp_group: "GroupCoordinator"
+    model: torch.nn.Module,
+    enable_compile: bool,
+    batch_size: int,
+    tp_group: "GroupCoordinator",
 ):
     """Patch the model to make it compatible with with torch.compile"""
     backup_ca_comm = None
 
     try:
         if enable_compile:
-            _to_torch(model)
+            _to_torch(model, reverse=False, batch_size=batch_size)
             monkey_patch_vllm_all_gather()
             backup_ca_comm = tp_group.ca_comm
             # Use custom-allreduce here.
             # We found the custom allreduce is much faster than the built-in allreduce in torch,
             # even with ENABLE_INTRA_NODE_COMM=1.
             # tp_group.ca_comm = None
             yield torch.compile(
-                torch.no_grad()(model.forward), mode="max-autotune-no-cudagraphs"
+                torch.no_grad()(model.forward),
+                mode="max-autotune-no-cudagraphs",
+                dynamic=False,
             )
         else:
             yield model.forward
     finally:
         if enable_compile:
-            _to_torch(model, reverse=True)
+            _to_torch(model, reverse=True, batch_size=batch_size)
             monkey_patch_vllm_all_gather(reverse=True)
             tp_group.ca_comm = backup_ca_comm
 
@@ -237,6 +245,7 @@ def capture(self):
                 with patch_model(
                     self.model_runner.model,
                     bs in self.compile_bs,
+                    bs,
                     self.model_runner.tp_group,
                 ) as forward:
                     (

python/sglang/srt/model_executor/model_runner.py

@@ -622,7 +622,7 @@ def init_cuda_graphs(self):
         tic = time.time()
         logger.info("Capture cuda graph begin. This can take up to several minutes.")
         self.cuda_graph_runner = CudaGraphRunner(self)
-        logger.info(f"Capture cuda graph end. Time elapsed: {time.time() - tic:.2f}s")
+        logger.info(f"Capture cuda graph end. Time elapsed: {time.time() - tic:.2f} s")
 
     def apply_torch_tp(self):
         logger.info(f"Enabling torch tensor parallelism on {self.tp_size} devices.")

test/srt/test_srt_engine.py

@@ -188,7 +188,7 @@ def test_8_engine_offline_throughput(self):
         )
         bench_args = BenchArgs(num_prompts=10)
         result = throughput_test(server_args=server_args, bench_args=bench_args)
-        self.assertGreater(result["total_throughput"], 3500)
+        self.assertGreater(result["total_throughput"], 3000)
 
 
 if __name__ == "__main__":

test/srt/test_torch_compile_moe.py

@@ -14,7 +14,7 @@
 )
 
 
-class TestTorchCompile(unittest.TestCase):
+class TestTorchCompileMoe(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = DEFAULT_SMALL_MOE_MODEL_NAME_FOR_TEST
@@ -23,7 +23,7 @@ def setUpClass(cls):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=["--enable-torch-compile", "--torch-compile-max-bs", "1"],
+            other_args=["--enable-torch-compile", "--torch-compile-max-bs", "8"],
         )
 
     @classmethod