[main] flashcomm_v1 before mlp weight prefetch in Qwen Dense Models

Signed-off-by: rjg-lyh <[email protected]>

Author: rjg-lyh

tests/e2e/multicard/test_offline_inference_distributed.py

@@ -23,13 +23,17 @@
 import os
 from unittest.mock import patch
 
+import pytest
+
 from modelscope import snapshot_download  # type: ignore
 from vllm import SamplingParams
 
 from tests.e2e.conftest import VllmRunner
 
 os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
 
+QWEN_DENSE_MODELS = ["Qwen/QwQ-32B", "Qwen/Qwen-32B"]
+
 
 def test_models_distributed_QwQ():
     example_prompts = [
@@ -150,3 +154,23 @@ def test_sp_for_qwen3_moe() -> None:
                     enable_expert_parallel=True,
                     enforce_eager=True) as vllm_model:
         vllm_model.generate(example_prompts, sampling_params)
+
+
+@pytest.mark.parametrize("enforce_eager", [True, False])
+@pytest.mark.parametrize("model", QWEN_DENSE_MODELS)
+@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE": "1"})
+@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_FLASHCOMM": "1"})
+def test_models_distributed_Qwen_Dense_with_flashcomm_v1(model, enforce_eager):
+    example_prompts = [
+        "Hello, my name is",
+    ]
+    max_tokens = 5
+
+    with VllmRunner(
+            snapshot_download(model),
+            max_model_len=8192,
+            enforce_eager=enforce_eager,
+            dtype="auto",
+            tensor_parallel_size=4,
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)

tests/ut/ops/test_layernorm.py

@@ -1,15 +1,24 @@
-from unittest.mock import patch
+import os
+from unittest.mock import patch, MagicMock
 
 import pytest
 import torch
 from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm_ascend.ascend_forward_context import set_ascend_forward_context
 
 
 @pytest.fixture
 def dummy_tensor():
     return torch.randn(4, 8, dtype=torch.float16)
 
 
+def mock_maybe_chunk_residual(x, residual):
+    if x.size(0) != residual.size(0):
+        return residual[:4]
+
+    return residual
+
+
 def mock_rms_norm(x, weight, eps):
     return x + 1, None
 
@@ -23,11 +32,12 @@ def mock_add_rms_norm(x, residual, weight, eps):
                          [None, torch.randn(4, 8, dtype=torch.float32)])
 @patch("torch_npu.npu_rms_norm", side_effect=mock_rms_norm)
 @patch("torch_npu.npu_add_rms_norm", side_effect=mock_add_rms_norm)
-def test_RMSNorm_forward(mock_add_rmsnorm, mock_rmsnorm, is_310p_return,
-                         residual, dummy_tensor):
+@patch("torch.ops.vllm.maybe_chunk_residual", side_effect=mock_maybe_chunk_residual)
+def test_RMSNorm_forward(mock_maybe_chunk_residual, mock_add_rmsnorm, mock_rmsnorm,
+                         is_310p_return,residual, dummy_tensor):
 
     with patch("vllm_ascend.utils.is_310p", return_value=is_310p_return):
-        layer = RMSNorm(hidden_size=32, eps=1e-05)
+        layer = RMSNorm(hidden_size=8, eps=1e-05)
         if residual is not None:
             out_x, out_residual = layer.forward_oot(dummy_tensor, residual)
 
@@ -51,3 +61,23 @@ def test_RMSNorm_forward(mock_add_rmsnorm, mock_rmsnorm, is_310p_return,
 
             mock_rmsnorm.assert_called_once()
             assert torch.allclose(out_x, expected_out_x)
+
+
+@patch("vllm_ascend.utils.is_310p", return_value=False)
+@patch("torch_npu.npu_add_rms_norm", side_effect=mock_add_rms_norm)
+@patch("torch.ops.vllm.maybe_chunk_residual", side_effect=mock_maybe_chunk_residual)
+def test_RMSNorm_forward_with_flashcomm_v1(mock_maybe_chunk_residual, mock_add_rms_norm, mock_is310p):
+    x = torch.randn(4, 512, dtype=torch.bfloat16)
+    residual = torch.randn(16, 512, dtype=torch.bfloat16)
+    layer = RMSNorm(hidden_size=512, eps=1e-05)
+
+    out_x, out_residual = layer.forward_oot(x, residual)
+
+    expected_out_x = 2 * x
+    expected_out_residual = 2 * residual[:4]
+
+    mock_maybe_chunk_residual.assert_called_once()
+    mock_add_rms_norm.assert_called_once()
+    assert out_residual.size(0) == 4
+    assert torch.allclose(out_x, expected_out_x)
+    assert torch.allclose(out_residual, expected_out_residual)

vllm_ascend/__init__.py

@@ -23,5 +23,10 @@ def register():
 
 
 def register_model():
+    import vllm.envs as envs
+
+    import vllm_ascend.envs as envs_ascend
+
     from .models import register_model
+
     register_model()

vllm_ascend/ascend_forward_context.py

@@ -105,6 +105,21 @@ def set_ascend_forward_context(
         # due to multiple warmups before actual capturing
         forward_context.capturing = False
 
+        # set this for rope forward_oot using
+        forward_context.is_first_layer = True
+
+        # set for flashcomm_v1
+        flashcomm_v1_enabled = envs_ascend.VLLM_ASCEND_ENABLE_FLASHCOMM and \
+            num_tokens is not None and num_tokens > 1000
+        
+        if flashcomm_v1_enabled:
+            tp_world_size = get_tensor_model_parallel_world_size()
+            pad_size = (tp_world_size -
+                        (num_tokens % tp_world_size)) % tp_world_size
+            forward_context.pad_size = pad_size
+
+        forward_context.flashcomm_v1_enabled = flashcomm_v1_enabled
+
         if num_tokens is None and attn_metadata is not None:
             num_tokens = attn_metadata.num_actual_tokens
 

vllm_ascend/envs.py

@@ -131,6 +131,13 @@
     # this feature is supported in A2, and eager mode will get better performance.
     "VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE":
     lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE", '0'))),
+    # Whether to enable FlashComm optimization when tensor parallel is enabled.
+    # this feature will get better performance in prefill phase.
+    "VLLM_ASCEND_ENABLE_FLASHCOMM":
+    lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_FLASHCOMM", '0'))),
+    # Whether to enable dense model and general optimizations for better performance.
+    "VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE":
+    lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE", '0'))),
     # Whether to enable mlp optimize when tensor parallel is enabled.
     # this feature in eager mode will get better performance.
     "VLLM_ASCEND_ENABLE_MLP_OPTIMIZE":

vllm_ascend/ops/__init__.py

@@ -24,6 +24,7 @@
 from vllm_ascend.ops.activation import AscendQuickGELU, AscendSiluAndMul
 from vllm_ascend.ops.rotary_embedding import (
     AscendDeepseekScalingRotaryEmbedding, AscendRotaryEmbedding)
+import vllm_ascend.ops.flashcomm_gate_ops
 
 
 class dummyFusionOp:

vllm_ascend/ops/flashcomm_gate_ops.py

@@ -0,0 +1,71 @@
+import torch
+import torch.nn.functional as F
+from vllm.utils import direct_register_custom_op
+from vllm.distributed import (tensor_model_parallel_all_gather,
+                              tensor_model_parallel_reduce_scatter,
+                              tensor_model_parallel_all_reduce,
+                              get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size)
+from vllm.forward_context import get_forward_context
+
+
+def _maybe_chunk_residual_impl(x: torch.Tensor, residual: torch.Tensor) -> torch.Tensor:
+    if x.size(0) != residual.size(0):
+        flashcomm_v1_enabled = get_forward_context().flashcomm_v1_enabled
+        assert flashcomm_v1_enabled is True, (
+            "Currently, this situation only occurs "
+            "when flashcomm_v1 is enabled"
+        )
+        tp_size = get_tensor_model_parallel_world_size()
+        tp_rank = get_tensor_model_parallel_rank()
+        residual = torch.chunk(residual, tp_size, dim=0)[tp_rank]
+
+    return residual
+
+
+def _maybe_all_gather_and_maybe_unpad_impl(x: torch.Tensor, label: bool) -> torch.Tensor:
+    flashcomm_v1_enabled = get_forward_context().flashcomm_v1_enabled
+    if flashcomm_v1_enabled and label:
+        x = tensor_model_parallel_all_gather(x, 0)
+        pad_size = get_forward_context().pad_size
+        if pad_size > 0:
+            x = x[:-pad_size, :]
+    return x
+
+
+def _maybe_pad_and_reduce_impl(x: torch.Tensor) -> torch.Tensor:
+    flashcomm_v1_enabled = get_forward_context().flashcomm_v1_enabled
+    if flashcomm_v1_enabled:
+        pad_size = get_forward_context().pad_size
+        if pad_size > 0:
+            x = F.pad(x, (0, 0, 0, pad_size))
+        return tensor_model_parallel_reduce_scatter(x, 0)
+    else:
+        return tensor_model_parallel_all_reduce(x)
+
+
+direct_register_custom_op(
+    op_name="maybe_chunk_residual",
+    op_func=_maybe_chunk_residual_impl,
+    fake_impl=lambda x, residual: residual,
+    mutates_args=[],
+    dispatch_key="PrivateUse1"
+)
+
+
+direct_register_custom_op(
+    op_name="maybe_all_gather_and_maybe_unpad",
+    op_func=_maybe_all_gather_and_maybe_unpad_impl,
+    fake_impl=lambda x, label: x,
+    mutates_args=[],
+    dispatch_key="PrivateUse1"
+)
+
+
+direct_register_custom_op(
+    op_name="maybe_pad_and_reduce",
+    op_func=_maybe_pad_and_reduce_impl,
+    fake_impl=lambda x: x,
+    mutates_args=[],
+    dispatch_key="PrivateUse1"
+)

vllm_ascend/ops/layernorm.py

@@ -44,6 +44,8 @@ def forward(
         import torch_npu
 
         if residual is not None:
+            residual = torch.ops.vllm.maybe_chunk_residual(x, residual)
+            assert x.size(0) == residual.size(0)
             x, _, residual = torch_npu.npu_add_rms_norm_quant(
                 x,
                 residual,
@@ -69,6 +71,8 @@ def forward_oot(
 
         from vllm_ascend.utils import is_310p
         if residual is not None:
+            residual = torch.ops.vllm.maybe_chunk_residual(x, residual)
+            assert x.size(0) == residual.size(0)
             if is_310p():
                 orig_dtype = residual.dtype
                 x = x + residual.to(x.dtype)

vllm_ascend/ops/linear.py

@@ -24,18 +24,22 @@
                               split_tensor_along_last_dim,
                               tensor_model_parallel_all_gather,
                               tensor_model_parallel_all_reduce)
-from vllm.model_executor.layers.linear import (WEIGHT_LOADER_V2_SUPPORTED,
-                                               ColumnParallelLinear,
-                                               LinearBase,
-                                               MergedColumnParallelLinear,
-                                               RowParallelLinear)
+from vllm.forward_context import get_forward_context
 from vllm.model_executor.layers.quantization.base_config import \
     QuantizationConfig
 from vllm.model_executor.utils import set_weight_attrs
 
 from vllm_ascend.distributed.parallel_state import (
     get_mlp_tensor_model_parallel_rank,
     get_mlp_tensor_model_parallel_world_size, get_mlp_tp_group)
+from vllm_ascend.quantization.w8a8 import AscendW8A8LinearMethod
+from vllm_ascend.utils import (all_gather_and_maybe_unpad,
+                               maybe_pad_and_reduce_scatter)
+
+from vllm.model_executor.layers.linear import (  # isort: skip
+    WEIGHT_LOADER_V2_SUPPORTED, ColumnParallelLinear, LinearBase,
+    MergedColumnParallelLinear, QKVParallelLinear, RowParallelLinear,
+    UnquantizedLinearMethod)
 
 
 class AscendMlpColumnParallelLinear(ColumnParallelLinear):
@@ -307,3 +311,103 @@ def forward(
         if not self.return_bias:
             return output
         return output, output_bias
+
+
+class AscendDenseMergedColumnParallelLinear(MergedColumnParallelLinear):
+    """Linear layer with column parallelism.
+
+    Implemented multiple optimization projects for dense models, such as FlashComm and
+    communication-computation fusion.
+    """
+
+    def forward(
+        self, input_: torch.Tensor
+    ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
+        bias = self.bias if not self.skip_bias_add else None
+
+        # Matrix multiply.
+        assert self.quant_method is not None
+
+        input_ = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(input_, True)
+        output_parallel = self.quant_method.apply(self, input_, bias)
+
+        if self.gather_output:
+            # All-gather across the partitions.
+            output = tensor_model_parallel_all_gather(output_parallel)
+        else:
+            output = output_parallel
+        output_bias = self.bias if self.skip_bias_add else None
+        if not self.return_bias:
+            return output
+        return output, output_bias
+
+
+class AscendDenseQKVParallelLinear(QKVParallelLinear):
+    """Linear layer with column parallelism.
+
+    Implemented multiple optimization projects for dense models, such as FlashComm and
+    communication-computation fusion.
+    """
+
+    def forward(
+        self, input_: torch.Tensor
+    ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
+        bias = self.bias if not self.skip_bias_add else None
+
+        # Matrix multiply.
+        assert self.quant_method is not None
+
+        layer_num = self.prefix.split('.')[2]
+
+        input_ = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(
+            input_, layer_num != '0')
+        output_parallel = self.quant_method.apply(self, input_, bias)
+
+        if self.gather_output:
+            # All-gather across the partitions.
+            output = tensor_model_parallel_all_gather(output_parallel)
+        else:
+            output = output_parallel
+        output_bias = self.bias if self.skip_bias_add else None
+        if not self.return_bias:
+            return output
+        return output, output_bias
+
+
+class AscendDenseRowParallelLinear(RowParallelLinear):
+    """Linear layer with row parallelism.
+
+    Implemented multiple optimization projects for dense models, such as FlashComm and
+    communication-computation fusion.
+    """
+
+    def forward(
+        self, input_: torch.Tensor
+    ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
+        if self.input_is_parallel:
+            input_parallel = input_
+        else:
+            tp_rank = get_tensor_model_parallel_rank()
+            splitted_input = split_tensor_along_last_dim(
+                input_, num_partitions=self.tp_size)
+            input_parallel = splitted_input[tp_rank].contiguous()
+
+        # Matrix multiply.
+        assert self.quant_method is not None 
+        # Only fuse bias add into GEMM for rank 0 (this ensures that
+        # bias will not get added more than once in TP>1 case)
+        bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
+        
+        if self.tp_size == 1 or not self.reduce_results:
+            output = self.quant_method.apply(self, input_parallel, bias=bias_)
+        else:
+            output_parallel = self.quant_method.apply(self,
+                                                      input_parallel,
+                                                      bias=bias_)
+            output = torch.ops.vllm.maybe_pad_and_reduce(output_parallel)
+
+        output_bias = self.bias if self.skip_bias_add else None
+
+        if not self.return_bias:
+            return output
+        return output, output_bias