update moe_sum and moe_align

vllm/envs.py

@@ -232,6 +232,9 @@ if TYPE_CHECKING:
     VLLM_USE_GLOBAL_CACHE13: bool = False
     VLLM_USE_LIGHTOP: bool = False
     VLLM_USE_OPT_CAT: bool = False
+    VLLM_USE_OPT_MOE_SUM: bool = False
+    VLLM_USE_LIGHTOP_MOE_SUM: bool = False
+    VLLM_USE_LIGHTOP_MOE_ALIGN: bool = False
     VLLM_USE_MERGE_ATTN_STATES_OPT: bool = False
     USE_FUSED_RMS_QUANT: bool = False
     USE_FUSED_SILU_MUL_QUANT: bool = False
@@ -1625,6 +1628,18 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_USE_OPT_CAT":
         lambda: (os.environ.get("VLLM_USE_OPT_CAT", "True").lower() in
                  ("true", "1")),  
+    # vLLM will use triton moe_sum 
+    "VLLM_USE_OPT_MOE_SUM":
+        lambda: (os.environ.get("VLLM_USE_OPT_MOE_SUM", "False").lower() in
+                 ("true", "1")),  
+    # vLLM will use lightop moe_sum 
+    "VLLM_USE_LIGHTOP_MOE_SUM":
+        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM", "False").lower() in
+                 ("true", "1")),  
+    # vLLM will use lightop moe_align_block_size 
+    "VLLM_USE_LIGHTOP_MOE_ALIGN":
+        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_ALIGN", "False").lower() in
+                 ("true", "1")),     
     # vLLM will use opt merge_aatn_states,not triton
     "VLLM_USE_MERGE_ATTN_STATES_OPT":
         lambda: (os.environ.get("VLLM_USE_MERGE_ATTN_STATES_OPT", "True").lower() in

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -58,6 +58,137 @@ logger = init_logger(__name__)
 
 if envs.VLLM_USE_GLOBAL_CACHE13:
     moe_cache_singleton = None
+    
+
+@torch.compile
+def moe_sum_reduce_torch_compile(x, out, routed_scaling_factor):
+    torch.sum(x, dim=1, out=out)
+    out.mul_(routed_scaling_factor)
+
+
+@triton.jit
+def _moe_sum_reduce_kernel(
+        input_ptr,
+        input_stride_0,
+        input_stride_1,
+        input_stride_2,
+        output_ptr,
+        output_stride_0,
+        output_stride_1,
+        token_num: int,
+        topk_num: int,
+        hidden_dim: int,
+        routed_scaling_factor: tl.constexpr,
+        BLOCK_M: tl.constexpr,
+        BLOCK_DIM: tl.constexpr,
+        NUM_STAGE: tl.constexpr,
+):
+    input_stride_0 = tl.cast(input_stride_0, dtype=tl.int64)
+    input_stride_1 = tl.cast(input_stride_1, dtype=tl.int64)
+    output_stride_0 = tl.cast(output_stride_0, dtype=tl.int64)
+
+    token_block_id = tl.program_id(0)
+    dim_block_id = tl.program_id(1)
+
+    token_start = token_block_id * BLOCK_M
+    token_end = min((token_block_id + 1) * BLOCK_M, token_num)
+
+    dim_start = dim_block_id * BLOCK_DIM
+    dim_end = min((dim_block_id + 1) * BLOCK_DIM, hidden_dim)
+
+    offs_dim = dim_start + tl.arange(0, BLOCK_DIM)
+
+    for token_index in range(token_start, token_end):
+        accumulator = tl.zeros((BLOCK_DIM,), dtype=tl.float32)
+        input_t_ptr = input_ptr + token_index * input_stride_0 + offs_dim
+        for i in tl.range(0, topk_num, num_stages=NUM_STAGE):
+            tmp = tl.load(
+                input_t_ptr + i * input_stride_1, mask=offs_dim < dim_end, other=0.0
+            )
+            accumulator += tmp
+        accumulator = accumulator * routed_scaling_factor
+        store_t_ptr = output_ptr + token_index * output_stride_0 + offs_dim
+        tl.store(
+            store_t_ptr,
+            accumulator.to(input_ptr.dtype.element_ty),
+            mask=offs_dim < dim_end,
+        )
+
+
+def moe_sum_reduce_triton(
+        input: torch.Tensor, output: torch.Tensor, routed_scaling_factor: float
+):
+    assert input.is_contiguous()
+    assert output.is_contiguous()
+
+    token_num, topk_num, hidden_dim = input.shape
+    assert output.shape[0] == token_num and output.shape[1] == hidden_dim
+
+    if token_num <= 32:
+        BLOCK_M = 1
+        BLOCK_DIM = 512
+        NUM_STAGE = 2
+        num_warps = 4
+
+    elif token_num <= 128:
+        BLOCK_M = 1
+        BLOCK_DIM = 1024
+        NUM_STAGE = 0
+        num_warps = 2
+
+    elif token_num <= 4096:
+        BLOCK_M = 1
+        BLOCK_DIM = 2048
+        NUM_STAGE = 0
+        num_warps = 2
+    else:
+        BLOCK_M = 1
+        BLOCK_DIM = 2048
+        NUM_STAGE = 2
+        num_warps = 8
+
+    grid = (
+        triton.cdiv(token_num, BLOCK_M),
+        triton.cdiv(hidden_dim, BLOCK_DIM),
+    )
+
+    _moe_sum_reduce_kernel[grid](
+        input,
+        *input.stride(),
+        output,
+        *output.stride(),
+        token_num=token_num,
+        topk_num=topk_num,
+        hidden_dim=hidden_dim,
+        routed_scaling_factor=routed_scaling_factor,
+        BLOCK_M=BLOCK_M,
+        BLOCK_DIM=BLOCK_DIM,
+        NUM_STAGE=NUM_STAGE,
+        num_warps=num_warps,
+    )
+    return
+
+
+def moe_reduce_dispatch(
+    intermediate_cache3: torch.Tensor,
+    out_hidden_states: torch.Tensor,
+    begin_chunk_idx: int,
+    end_chunk_idx: int,
+):
+    inter_cache_view = intermediate_cache3.view(*intermediate_cache3.shape)
+    n = intermediate_cache3.shape[0]
+
+    # 根据 n 大小选择不同的 reduce 实现
+    if 1 <= n <= 4:
+        moe_sum_reduce_torch_compile(inter_cache_view, out_hidden_states[begin_chunk_idx:end_chunk_idx], 1.0)
+    elif 4 < n <= 1024:
+        moe_sum_reduce_triton(inter_cache_view, out_hidden_states[begin_chunk_idx:end_chunk_idx], 1.0)
+    elif 1024 < n <= 32768:
+        ops.moe_sum_opt1(inter_cache_view, out_hidden_states[begin_chunk_idx:end_chunk_idx])
+    else:
+        ops.moe_sum(inter_cache_view, out_hidden_states[begin_chunk_idx:end_chunk_idx])
+
+
 def get_moe_cache(top_k_num,N,K,device,dtype):
     global moe_cache_singleton
     if moe_cache_singleton is None:
@@ -2046,6 +2177,14 @@ def fused_experts_impl(
                                 B_bias=w2_bias,
                                 use_nn_moe=use_nn_moe)
 
+        if envs.VLLM_USE_LIGHTOP_MOE_SUM:
+                from lightop import op as op
+                op.moe_sum(input=intermediate_cache3.view(*intermediate_cache3.size()),
+                    output=out_hidden_states[begin_chunk_idx:end_chunk_idx], bias=None, 
+                    expert_mask=None, num_local_tokens=None, factor=1.0)
+        elif envs.VLLM_USE_OPT_MOE_SUM:
+                moe_reduce_dispatch(intermediate_cache3.view(*intermediate_cache3.size()), out_hidden_states[begin_chunk_idx:end_chunk_idx], begin_chunk_idx, end_chunk_idx)
+        else:
             ops.moe_sum(intermediate_cache3.view(*intermediate_cache3.size()),
                             out_hidden_states[begin_chunk_idx:end_chunk_idx])
 

vllm/model_executor/layers/fused_moe/moe_align_block_size.py

@@ -102,6 +102,14 @@ def moe_align_block_size(
                                 expert_map = expert_map,
                                 expert_mask = expert_mask,
                                 num_local_tokens = None)
+    else:
+        if envs.VLLM_USE_LIGHTOP_MOE_ALIGN:
+            from lightop import op as op
+            op.moe_align_block_size(topk_ids, num_experts, block_size, sorted_ids,
+                                expert_ids, num_tokens_post_pad,
+                                expert_map = None,
+                                expert_mask = None,
+                                num_local_tokens = None)
         else:
             ops.moe_align_block_size(topk_ids, num_experts, block_size, sorted_ids,
                                     expert_ids, num_tokens_post_pad)