Add Triton MoE Kernel

aiter/ops/triton/moe_op.py

@@ -0,0 +1,278 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+import torch
+import triton
+import triton.language as tl
+from typing import Any, Dict, Optional, List
+
+
+#TODO: add support to set this through env var or a function call by a framework like VLLM/SGLANG
+#MOE_PADDING_SIZE = 128 if AITER_TRITON_MOE_PADDING else 0
+MOE_PADDING_SIZE = 0
+
+@triton.jit
+def _fused_moe_kernel(
+        # Pointers to matrices
+        a_ptr,
+        b_ptr,
+        c_ptr,
+        a_scale_ptr,
+        b_scale_ptr,
+        topk_weights_ptr,
+        sorted_token_ids_ptr,
+        expert_ids_ptr,
+        num_tokens_post_padded_ptr,
+        # Matrix dimensions
+        N,
+        K,
+        EM,
+        num_valid_tokens,
+        # The stride variables represent how much to increase the ptr by when
+        # moving by 1 element in a particular dimension. E.g. `stride_am` is
+        # how much to increase `a_ptr` by to get the element one row down
+        # (A has M rows).
+        stride_am,
+        stride_ak,
+        stride_be,
+        stride_bk,
+        stride_bn,
+        stride_cm,
+        stride_cn,
+        stride_asm,
+        stride_ask,
+        stride_bse,
+        stride_bsk,
+        stride_bsn,
+        # Block size for block-wise quantization
+        group_n: tl.constexpr,
+        group_k: tl.constexpr,
+        # Meta-parameters
+        BLOCK_SIZE_M: tl.constexpr,
+        BLOCK_SIZE_N: tl.constexpr,
+        BLOCK_SIZE_K: tl.constexpr,
+        GROUP_SIZE_M: tl.constexpr,
+        MUL_ROUTED_WEIGHT: tl.constexpr,
+        top_k: tl.constexpr,
+        compute_type: tl.constexpr,
+        use_fp8_w8a8: tl.constexpr,
+        use_int8_w8a16: tl.constexpr):
+    """
+    Note: this is Triton jited function and not meant to be called directly. Call fused_moe function
+    below
+
+    Implements the fused computation for a Mixture of Experts (MOE) using
+    token and expert matrices.
+
+
+    Key Parameters:
+    - A: The input tensor representing tokens with shape (*, K), where '*' can
+        be any shape representing batches and K is the feature dimension of
+        each token.
+    - B: The stacked MOE weight tensor with shape (E, N, K), where E is
+        the number of experts, K is the input feature dimension, and N is
+        the output feature dimension.
+    - C: The output cache tensor with shape (M, topk, N), where M is the
+        total number of tokens post padding, topk is the number of times
+        each token is repeated, and N is the output feature dimension.
+    - sorted_token_ids: A tensor containing the sorted indices of tokens,
+        repeated topk times and arranged by the expert index they are
+        assigned to.
+    - expert_ids: A tensor containing the indices of the expert for each
+        block. It determines which expert matrix from B should be used for
+        each block in A.
+    This kernel performs the multiplication of a token by its corresponding
+    expert matrix as determined by `expert_ids`. The sorting of
+    `sorted_token_ids` by expert index and padding ensures divisibility by
+    BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
+    multiplication across different blocks processed by the same expert.
+    """
+    # -----------------------------------------------------------
+    # Map program ids `pid` to the block of C it should compute.
+    # This is done in a grouped ordering to promote L2 data reuse.
+    pid = tl.program_id(axis=0)
+    num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
+    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
+    num_pid_in_group = GROUP_SIZE_M * num_pid_n
+    group_id = pid // num_pid_in_group
+    first_pid_m = group_id * GROUP_SIZE_M
+    group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
+    pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
+    pid_n = (pid % num_pid_in_group) // group_size_m
+
+    # ----------------------------------------------------------
+    # Create pointers for the first blocks of A and B.
+    # We will advance this pointer as we move in the K direction
+    # and accumulate
+    # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
+    # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
+    num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
+    if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
+        return
+    offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(
+        tl.int64)
+    offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
+    token_mask = offs_token < num_valid_tokens
+
+    offs_bn = (pid_n * BLOCK_SIZE_N +
+               tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
+    offs_k = tl.arange(0, BLOCK_SIZE_K)
+    a_ptrs = a_ptr + (offs_token[:, None] // top_k * stride_am +
+                      offs_k[None, :] * stride_ak)
+
+    off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
+    b_ptrs = b_ptr + off_experts * stride_be + (offs_k[:, None] * stride_bk +
+                                                offs_bn[None, :] * stride_bn)
+    if use_int8_w8a16:
+        b_scale_ptrs = b_scale_ptr + off_experts * stride_bse + offs_bn[
+            None, :] * stride_bsn
+        b_scale = tl.load(b_scale_ptrs)
+
+    if use_fp8_w8a8:
+        if group_k > 0 and group_n > 0:
+            a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
+            offs_bsn = offs_bn // group_n
+            b_scale_ptrs = (b_scale_ptr + off_experts * stride_bse +
+                            offs_bsn * stride_bsn)
+        else:
+            a_scale = tl.load(a_scale_ptr)
+            b_scale = tl.load(b_scale_ptr + off_experts)
+
+    # -----------------------------------------------------------
+    # Iterate to compute a block of the C matrix.
+    # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
+    # of fp32 values for higher accuracy.
+    # `accumulator` will be converted back to fp16 after the loop.
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+
+    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
+        # Load the next block of A and B, generate a mask by checking the
+        # K dimension.
+        a = tl.load(a_ptrs,
+                    mask=token_mask[:, None] &
+                    (offs_k[None, :] < K - k * BLOCK_SIZE_K),
+                    other=0.0)
+        b = tl.load(b_ptrs,
+                    mask=offs_k[:, None] < K - k * BLOCK_SIZE_K,
+                    other=0.0)
+        # We accumulate along the K dimension.
+        if use_int8_w8a16:
+            accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
+        elif use_fp8_w8a8:
+            if group_k > 0 and group_n > 0:
+                k_start = k * BLOCK_SIZE_K
+                offs_ks = k_start // group_k
+                a_scale = tl.load(a_scale_ptrs + offs_ks * stride_ask,
+                                  mask=token_mask,
+                                  other=0.0)
+                b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk)
+
+                accumulator += tl.dot(a, b) * a_scale[:,
+                                                      None] * b_scale[None, :]
+            else:
+                accumulator = tl.dot(a, b, acc=accumulator)
+        else:
+            accumulator += tl.dot(a, b)
+        # Advance the ptrs to the next K block.
+        a_ptrs += BLOCK_SIZE_K * stride_ak
+        b_ptrs += BLOCK_SIZE_K * stride_bk
+
+    if MUL_ROUTED_WEIGHT:
+        moe_weight = tl.load(topk_weights_ptr + offs_token,
+                             mask=token_mask,
+                             other=0)
+        accumulator = accumulator * moe_weight[:, None]
+    if use_int8_w8a16:
+        accumulator = (accumulator * b_scale).to(compute_type)
+    elif use_fp8_w8a8:
+        if group_k > 0 and group_n > 0:
+            accumulator = accumulator.to(compute_type)
+        else:
+            accumulator = (accumulator * a_scale * b_scale).to(compute_type)
+    else:
+        accumulator = accumulator.to(compute_type)
+    # -----------------------------------------------------------
+    # Write back the block of the output
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[
+        None, :]
+    c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
+    tl.store(c_ptrs, accumulator, mask=c_mask)
+
+def moe_triton(A: torch.Tensor,
+                            B: torch.Tensor,
+                            C: torch.Tensor,
+                            A_scale: Optional[torch.Tensor],
+                            B_scale: Optional[torch.Tensor],
+                            topk_weights: torch.Tensor,
+                            topk_ids: torch.Tensor,
+                            sorted_token_ids: torch.Tensor,
+                            expert_ids: torch.Tensor,
+                            num_tokens_post_padded: torch.Tensor,
+                            mul_routed_weight: bool,
+                            top_k: int,
+                            config: Dict[str, Any],
+                            compute_type: tl.dtype,
+                            use_fp8_w8a8: bool,
+                            use_int8_w8a16: bool,
+                            block_shape: Optional[List[int]] = None) -> None:
+    if mul_routed_weight:
+        assert topk_weights.stride(1) == 1
+    assert sorted_token_ids.stride(0) == 1
+
+    if use_fp8_w8a8:
+        assert B_scale is not None
+        if block_shape is None:
+            A, A_scale = ops.scaled_fp8_quant(A, A_scale)
+        else:
+            assert len(block_shape) == 2
+            block_n, block_k = block_shape[0], block_shape[1]
+            A, A_scale = per_token_group_quant_fp8(A, block_k)
+            assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
+            assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
+            assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
+    elif use_int8_w8a16:
+        assert B_scale is not None
+    else:
+        assert A_scale is None
+        assert B_scale is None
+
+    grid = lambda META: (triton.cdiv(sorted_token_ids.shape[0], META[
+        'BLOCK_SIZE_M']) * triton.cdiv(B.shape[1], META['BLOCK_SIZE_N']), )
+
+    _fused_moe_kernel[grid](
+        A,
+        B,
+        C,
+        A_scale,
+        B_scale,
+        topk_weights,
+        sorted_token_ids,
+        expert_ids,
+        num_tokens_post_padded,
+        B.shape[1],
+        B.shape[2] - MOE_PADDING_SIZE,
+        sorted_token_ids.shape[0],
+        topk_ids.numel(),
+        A.stride(0),
+        A.stride(1),
+        B.stride(0),
+        B.stride(2),
+        B.stride(1),
+        C.stride(1),
+        C.stride(2),
+        A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
+        A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
+        B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
+        B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
+        B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
+        0 if block_shape is None else block_shape[0],
+        0 if block_shape is None else block_shape[1],
+        MUL_ROUTED_WEIGHT=mul_routed_weight,
+        top_k=top_k,
+        compute_type=compute_type,
+        use_fp8_w8a8=use_fp8_w8a8,
+        use_int8_w8a16=use_int8_w8a16,
+        **config,
+    )
+