Add async tp pass

cascade812 · cascade812 · commit 13cd0bde7a99 · 2025-05-11T04:55:45.000Z
Signed-off-by: cascade812 &lt;cascade812@outlook.com&gt;
diff --git a/.buildkite/test-pipeline.yaml b/.buildkite/test-pipeline.yaml
@@ -309,6 +309,7 @@ steps:
     - pytest -v -s compile/test_pass_manager.py
     - pytest -v -s compile/test_fusion.py
     - pytest -v -s compile/test_sequence_parallelism.py
+    - pytest -v -s compile/test_async_tp.py
 
 - label: PyTorch Fullgraph Smoke Test # 9min
   mirror_hardwares: [amdexperimental, amdproduction]
diff --git a/tests/compile/test_async_tp.py b/tests/compile/test_async_tp.py
@@ -0,0 +1,186 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import pytest
+import torch
+
+import vllm.envs as envs
+from vllm.compilation.collective_fusion import AsyncTPPass
+from vllm.compilation.fx_utils import (find_specified_fn,
+                                       find_specified_fn_maybe)
+from vllm.config import (CompilationConfig, DeviceConfig, ModelConfig,
+                         PassConfig, VllmConfig)
+from vllm.distributed import (tensor_model_parallel_all_gather,
+                              tensor_model_parallel_reduce_scatter)
+from vllm.distributed.parallel_state import (init_distributed_environment,
+                                             initialize_model_parallel)
+from vllm.platforms import current_platform
+from vllm.utils import update_environment_variables
+
+from ..utils import multi_gpu_test
+from .backend import TestBackend
+
+prompts = [
+    "Hello, my name is",
+    "The president of the United States is",
+    "The capital of France is",
+    "The future of AI is",
+]
+
+
+class TestMMRSModel(torch.nn.Module):
+
+    def __init__(self, hidden_size=16):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.gate_proj = torch.nn.Parameter(torch.empty(
+            (self.hidden_size * 2, hidden_size)),
+                                            requires_grad=False)
+        # Initialize weights
+        torch.nn.init.normal_(self.gate_proj, std=0.02)
+
+    def forward(self, hidden_states):
+        """
+        Forward pass implementing the mm + reduce scatter in the FX graph
+    
+        """
+        # Reshape input
+        view = hidden_states.reshape(-1, self.hidden_size)
+
+        # matrix multiplication
+        permute = self.gate_proj.permute(1, 0)
+        mm = torch.mm(view, permute)
+        reduce_scatter = tensor_model_parallel_reduce_scatter(mm, dim=0)
+        return reduce_scatter
+
+    def ops_in_model_before(self):
+        return [torch.ops.vllm.reduce_scatter.default]
+
+    def ops_in_model_after(self):
+        return [torch.ops.symm_mem.fused_matmul_reduce_scatter.default]
+
+
+class TestAGMMModel(torch.nn.Module):
+
+    def __init__(self, hidden_size=16):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.weight = torch.nn.Parameter(torch.empty(
+            (hidden_size, hidden_size)),
+                                         requires_grad=False)
+        # Initialize weights
+        torch.nn.init.normal_(self.weight, std=0.02)
+
+    def forward(self, hidden_states):
+        """
+        Forward pass implementing the mm + all gather in the FX graph
+        """
+        # Reshape input
+        view = hidden_states.reshape(-1, self.hidden_size)
+        all_gather = tensor_model_parallel_all_gather(view, dim=0)
+        permute = self.weight.permute(1, 0)
+        mm = torch.mm(all_gather, permute)
+        return mm
+
+    def ops_in_model_before(self):
+        return [torch.ops.vllm.all_gather.default]
+
+    def ops_in_model_after(self):
+        return [torch.ops.symm_mem.fused_all_gather_matmul.default]
+
+
+@multi_gpu_test(num_gpus=2)
+@pytest.mark.parametrize("test_model", ["TestMMRSModel", "TestAGMMModel"])
+@pytest.mark.parametrize("batch_size", [8])
+@pytest.mark.parametrize("seq_len", [16])
+@pytest.mark.parametrize("hidden_size", [16])
+@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
+@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE not in ["cuda"],
+                    reason="Only test on CUDA")
+def test_sequence_parallelism_pass(test_model: str, batch_size: int,
+                                   seq_len: int, hidden_size: int,
+                                   dtype: torch.dtype):
+    num_processes = 2
+
+    def run_torch_spawn(fn, nprocs):
+        # need to use torch.mp.spawn otherwise will have problems with
+        # torch.distributed and cuda
+        torch.multiprocessing.spawn(fn,
+                                    args=(num_processes, test_model,
+                                          batch_size, seq_len, hidden_size,
+                                          dtype),
+                                    nprocs=nprocs)
+
+    run_torch_spawn(async_tp_pass_on_test_model, num_processes)
+
+
+def async_tp_pass_on_test_model(local_rank: int, world_size: int,
+                                test_model: str, batch_size: int, seq_len: int,
+                                hidden_size: int, dtype: torch.dtype):
+    current_platform.seed_everything(0)
+
+    device = torch.device(f"cuda:{local_rank}")
+    torch.cuda.set_device(device)
+    torch.set_default_device(device)
+    torch.set_default_dtype(dtype)
+
+    update_environment_variables({
+        'RANK': str(local_rank),
+        'LOCAL_RANK': str(local_rank),
+        'WORLD_SIZE': str(world_size),
+        'MASTER_ADDR': 'localhost',
+        'MASTER_PORT': '12345',
+    })
+
+    # initialize distributed
+    init_distributed_environment()
+    initialize_model_parallel(tensor_model_parallel_size=world_size)
+
+    # configure vllm config for SequenceParallelismPass
+    vllm_config = VllmConfig()
+    vllm_config.compilation_config = CompilationConfig(pass_config=PassConfig(
+        enable_async_tp=True, ), )
+    vllm_config.device_config = DeviceConfig(device=torch.device("cuda"))
+
+    # this is a fake model name to construct the model config
+    # in the vllm_config, it's not really used.
+    model = "nm-testing/TinyLlama-1.1B-Chat-v1.0-FP8-e2e"
+    vllm_config.model_config = ModelConfig(model=model,
+                                           task="auto",
+                                           tokenizer=model,
+                                           tokenizer_mode="auto",
+                                           trust_remote_code=True,
+                                           dtype=dtype,
+                                           seed=42)
+
+    async_tp_pass = AsyncTPPass(vllm_config)
+    backend = TestBackend(async_tp_pass)
+
+    if test_model == "TestMMRSModel":
+        model = TestMMRSModel(hidden_size)
+    elif test_model == "TestAGMMModel":
+        model = TestAGMMModel(hidden_size)
+    else:
+        raise ValueError(f"Unknown model: {test_model}")
+
+    hidden_states = torch.randn((batch_size * seq_len, hidden_size),
+                                dtype=dtype,
+                                requires_grad=False)
+
+    compiled_model = torch.compile(model, backend=backend)
+    compiled_model(hidden_states)
+
+    # Check substitution worked
+    pre_nodes = backend.graph_pre_pass.nodes
+    post_nodes = backend.graph_post_pass.nodes
+
+    # In pre-nodes, all reduce should exist,
+    # fused_matmul_reduce_scatter or fused_all_gather_matmul should not
+    for op in model.ops_in_model_before():
+        find_specified_fn(pre_nodes, op)
+    for op in model.ops_in_model_after():
+        assert find_specified_fn_maybe(pre_nodes, op) is None
+
+    # In post-nodes, fused_matmul_reduce_scatter or \
+    # fused_all_gather_matmul should exist
+    for op in model.ops_in_model_after():
+        find_specified_fn(post_nodes, op)
diff --git a/vllm/compilation/collective_fusion.py b/vllm/compilation/collective_fusion.py
@@ -0,0 +1,125 @@
+# SPDX-License-Identifier: Apache-2.0
+from typing import Optional, Tuple
+
+import torch
+import torch._inductor.pattern_matcher as pm
+import torch.fx as fx
+from torch._inductor.pattern_matcher import PatternMatcherPass
+from torch.distributed._symmetric_memory import enable_symm_mem_for_group
+
+from vllm.config import VllmConfig
+from vllm.distributed import get_tp_group
+from vllm.distributed.parallel_state import (
+    get_tensor_model_parallel_world_size)
+from vllm.logger import init_logger
+
+from .vllm_inductor_pass import VllmInductorPass
+
+logger = init_logger(__name__)
+
+
+class BasePattern:
+
+    def __init__(self, dtype: torch.dtype, device: str):
+        self.dtype = dtype
+        self.device = device
+        self.tp = get_tp_group()
+        self.tp_size = get_tensor_model_parallel_world_size()
+
+
+class GEMMReduceScatterPattern(BasePattern):
+
+    def get_inputs(self):
+        mul = torch.empty([16, 4], device=self.device, dtype=self.dtype)
+        mm_weight = torch.empty([4, 4], device=self.device, dtype=self.dtype)
+        return [mul, mm_weight]
+
+    def register(self, pm_pass: PatternMatcherPass):
+
+        def pattern(mul: torch.Tensor, mm_weight: torch.Tensor):
+            mm = torch.ops.aten.mm.default(mul, mm_weight)
+            reduce_scatter = torch.ops.vllm.reduce_scatter.default(
+                mm,
+                dim=0,
+                world_size=self.tp_size,
+                group_name=self.tp.unique_name)
+            return reduce_scatter
+
+        def replacement(mul: torch.Tensor, mm_weight: torch.Tensor):
+            gemm_rs = torch.ops.symm_mem.fused_matmul_reduce_scatter(
+                mul,
+                mm_weight,
+                "avg",
+                scatter_dim=0,
+                group_name=self.tp.device_group.group_name,
+            )
+
+            return gemm_rs
+
+        pm.register_replacement(pattern, replacement, self.get_inputs(),
+                                pm.fwd_only, pm_pass)
+
+
+class AllGatherGEMMPattern(BasePattern):
+
+    def get_inputs(self):
+        x = torch.empty([4, 4], device=self.device, dtype=self.dtype)
+        weight = torch.empty([4, 4], device=self.device, dtype=self.dtype)
+
+        return [x, weight]
+
+    def register(self, pm_pass: PatternMatcherPass):
+
+        def pattern(
+            x: torch.Tensor,
+            weight: torch.Tensor,
+        ) -> Tuple[torch.Tensor, torch.Tensor]:
+            all_gather = torch.ops.vllm.all_gather.default(
+                x,
+                dim=0,
+                world_size=self.tp_size,
+                group_name=self.tp.unique_name)
+
+            return torch.ops.aten.mm.default(all_gather, weight)
+
+        def replacement(
+                x: torch.Tensor,
+                weight: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+            ag_output, mm_outputs = torch.ops.symm_mem.fused_all_gather_matmul(
+                x,
+                [weight],
+                gather_dim=0,
+                group_name=self.tp.device_group.group_name,
+            )
+            return mm_outputs
+
+        pm.register_replacement(pattern, replacement, self.get_inputs(),
+                                pm.fwd_only, pm_pass)
+
+
+class AsyncTPPass(VllmInductorPass):
+
+    def __init__(self, config: VllmConfig):
+        super().__init__(config)
+
+        # Enable symmetric memory for the TP process group
+        enable_symm_mem_for_group(get_tp_group().device_group.group_name)
+        self.patterns: PatternMatcherPass = PatternMatcherPass(
+            pass_name="async_tp_pass")
+        GEMMReduceScatterPattern(self.dtype,
+                                 self.device).register(self.patterns)
+
+        AllGatherGEMMPattern(self.dtype, self.device).register(self.patterns)
+
+    def is_applicable_for_shape(self, shape: Optional[int]) -> bool:
+        # only do replace for specific shapes
+        tp_size = get_tensor_model_parallel_world_size()
+        return shape is not None and shape % tp_size == 0
+
+    def __call__(self, graph: fx.Graph):
+        self.begin()
+        self.dump_graph(graph, "before_async_tp_pass")
+        count = self.patterns.apply(graph)
+        logger.debug("Replaced %s patterns", count)
+        self.dump_graph(graph, "after_async_tp_pass")
+        self.end_and_log()
diff --git a/vllm/compilation/pass_manager.py b/vllm/compilation/pass_manager.py
@@ -8,6 +8,7 @@
 from vllm.logger import init_logger
 
 from .activation_quant_fusion import ActivationQuantFusionPass
+from .collective_fusion import AsyncTPPass
 from .fix_functionalization import FixFunctionalizationPass
 from .fusion import FusionPass
 from .inductor_pass import CustomGraphPass, InductorPass, get_pass_context
@@ -56,6 +57,8 @@ def configure(self, config: VllmConfig):
 
         if self.pass_config.enable_sequence_parallelism:
             self.passes += [SequenceParallelismPass(config)]
+            if self.pass_config.enable_async_tp:
+                self.passes += [AsyncTPPass(config)]
 
         self.fix_functionalization = FixFunctionalizationPass(config)
 
diff --git a/vllm/compilation/sequence_parallelism.py b/vllm/compilation/sequence_parallelism.py
@@ -255,12 +255,13 @@ def __init__(self, config: VllmConfig):
             torch._inductor.pattern_matcher._seen_patterns.clear()
 
     def is_applicable_for_shape(self, shape: Optional[int]) -> bool:
-        # only do replace for specific shapes
         tp_size = get_tensor_model_parallel_world_size()
         return shape is not None and shape % tp_size == 0
 
     def __call__(self, graph: fx.Graph):
+        self.begin()
         self.dump_graph(graph, "before_sequence_parallelism_pass")
         count = self.patterns.apply(graph)
         logger.debug("Replaced %s patterns", count)
         self.dump_graph(graph, "after_sequence_parallelism_pass")
+        self.end_and_log()
diff --git a/vllm/config.py b/vllm/config.py
diff --git a/vllm/distributed/parallel_state.py b/vllm/distributed/parallel_state.py
