dsv4.1: Hopper FP8 matmul kernels and tuning (#39657)

Co-authored-by: BBuf <1182563586@qq.com>
Co-authored-by: Yuhao Yang <47235274+yhyang201@users.noreply.github.com>

python/sglang/kernels/ops/quantization/configs/N=1152,K=5120,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[32, 32].json

@@ -0,0 +1,20 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4,
+        "SWAP_AB": true,
+        "SPLIT_K": 16
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/kernels/ops/quantization/configs/N=1792,K=5120,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[32, 32].json

@@ -0,0 +1,20 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4,
+        "SWAP_AB": true,
+        "SPLIT_K": 8
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/kernels/ops/quantization/configs/N=25600,K=6144,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[32, 32].json

@@ -0,0 +1,19 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4,
+        "SWAP_AB": true
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/kernels/ops/quantization/configs/N=4096,K=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[32, 32].json

@@ -0,0 +1,20 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4,
+        "SWAP_AB": true,
+        "SPLIT_K": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/kernels/ops/quantization/configs/N=5120,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[32, 32].json

@@ -0,0 +1,20 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4,
+        "SWAP_AB": true,
+        "SPLIT_K": 8
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/kernels/ops/quantization/configs/N=5120,K=576,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[32, 32].json

@@ -0,0 +1,20 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4,
+        "SWAP_AB": true,
+        "SPLIT_K": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/kernels/ops/quantization/configs/N=8192,K=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[32, 32].json

@@ -0,0 +1,20 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4,
+        "SWAP_AB": true,
+        "SPLIT_K": 2
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 32,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/kernels/ops/quantization/fp8_kernel.py

@@ -26,6 +26,7 @@ import triton.language as tl
 from sglang.kernels.jit.utils import is_arch_support_pdl
 from sglang.kernels.ops.quantization.fp8_utils import fp8_dtype_to_triton
 from sglang.srt.layers import deep_gemm_wrapper
+from sglang.srt.runtime_context import get_platform
 from sglang.srt.utils import (
     ceil_align,
     get_bool_env_var,
@@ -1029,6 +1030,133 @@ def _w8a8_block_fp8_matmul(
     tl.store(c_ptrs, c, mask=c_mask)
 
 
+@triton.jit
+def _w8a8_block_fp8_matmul_hopper(
+    # Pointers to inputs and output
+    A,
+    B,
+    C,
+    As,
+    Bs,
+    # Shape for matmul
+    M,
+    N,
+    K,
+    # Block size for block-wise quantization
+    group_n,
+    group_k,
+    # Stride for inputs and output
+    stride_am,
+    stride_ak,
+    stride_bk,
+    stride_bn,
+    stride_cm,
+    stride_cn,
+    stride_As_m,
+    stride_As_k,
+    stride_Bs_k,
+    stride_Bs_n,
+    # Meta-parameters
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_N: tl.constexpr,
+    BLOCK_SIZE_K: tl.constexpr,
+    GROUP_SIZE_M: tl.constexpr,
+    needs_masking: tl.constexpr,
+    SWAP_AB: tl.constexpr = False,
+    SPLIT_K: tl.constexpr = 1,
+):
+
+    pid = tl.program_id(axis=0)
+    split = tl.program_id(axis=1)
+    tiles_per_split = tl.cdiv(tl.cdiv(K, BLOCK_SIZE_K), SPLIT_K)
+    first_tile = split * tiles_per_split
+    C += split * M * N
+    num_pid_m = tl.cdiv(M, 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 % group_size_m)
+    pid_n = (pid % num_pid_in_group) // group_size_m
+
+    offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
+    offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
+    offs_k = tl.arange(0, BLOCK_SIZE_K)
+    a_ptrs = A + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak)
+    b_ptrs = B + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)
+
+    As_ptrs = As + offs_am * stride_As_m
+    offs_bsn = offs_bn // group_n
+    Bs_ptrs = Bs + offs_bsn * stride_Bs_n
+    n_tiles_k_per_group_k = group_k // BLOCK_SIZE_K
+
+    a_ptrs += first_tile * BLOCK_SIZE_K * stride_ak
+    b_ptrs += first_tile * BLOCK_SIZE_K * stride_bk
+    As_ptrs += (first_tile // n_tiles_k_per_group_k) * stride_As_k
+    Bs_ptrs += (first_tile // n_tiles_k_per_group_k) * stride_Bs_k
+
+    # Small-M Hopper configs transpose the MMA so the weight tile occupies M.
+    if SWAP_AB:
+        accumulator = tl.zeros((BLOCK_SIZE_N, BLOCK_SIZE_M), dtype=tl.float32)
+    else:
+        accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+    for k in range(
+        first_tile, tl.minimum(first_tile + tiles_per_split, tl.cdiv(K, BLOCK_SIZE_K))
+    ):
+        if needs_masking:
+            a = tl.load(a_ptrs, mask=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)
+        else:
+            a = tl.load(a_ptrs)
+            b = tl.load(b_ptrs)
+
+        a_s = tl.load(As_ptrs)
+        b_s = tl.load(Bs_ptrs)
+
+        scale_step_k = tl.where((k + 1) % n_tiles_k_per_group_k == 0, 1, 0)
+        if SWAP_AB:
+            accumulator += (
+                tl.dot(tl.trans(b), tl.trans(a)) * b_s[:, None] * a_s[None, :]
+            )
+        else:
+            accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :]
+        a_ptrs += BLOCK_SIZE_K * stride_ak
+        b_ptrs += BLOCK_SIZE_K * stride_bk
+        As_ptrs += scale_step_k * stride_As_k
+        Bs_ptrs += scale_step_k * stride_Bs_k
+
+    if SWAP_AB:
+        accumulator = tl.trans(accumulator)
+
+    if C.dtype.element_ty == tl.bfloat16:
+        c = accumulator.to(tl.bfloat16)
+    elif C.dtype.element_ty == tl.float16:
+        c = accumulator.to(tl.float16)
+    else:
+        c = accumulator.to(tl.float32)
+
+    offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = C + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
+    c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
+    tl.store(c_ptrs, c, mask=c_mask)
+
+
+@triton.jit
+def _reduce_block_fp8_split_k(
+    Parts, Out, ELEMENTS: tl.constexpr, SPLITS: tl.constexpr, BLOCK: tl.constexpr
+):
+    offsets = tl.program_id(0) * BLOCK + tl.arange(0, BLOCK)
+    splits = tl.arange(0, SPLITS)
+    values = tl.load(
+        Parts + splits[:, None] * ELEMENTS + offsets[None, :],
+        offsets[None, :] < ELEMENTS,
+        0.0,
+    )
+    tl.store(Out + offsets, tl.sum(values, axis=0), offsets < ELEMENTS)
+
+
 @triton.jit
 def _w8a8_block_fp8_matmul_gfx1250(
     # Pointers to inputs and output
@@ -1579,23 +1707,38 @@ def w8a8_block_fp8_matmul_triton(
             "num_stages": 3,
         }
 
-    if _is_gfx1250:
+    # Split-K accumulates K in SPLIT_K separate fp32 partials, so its results
+    # do not match the single-accumulator kernels bit-for-bit.
+    hopper_tuned = get_platform().is_sm90 and (
+        config.get("SWAP_AB", False) or config.get("SPLIT_K", 1) > 1
+    )
+    if hopper_tuned:
+        kernel = _w8a8_block_fp8_matmul_hopper
+    elif _is_gfx1250:
         config = {**config, "num_stages": 1}
         kernel = _w8a8_block_fp8_matmul_gfx1250
     else:
         kernel = select_w8a8_block_fp8_matmul_kernel(M, N, config)
 
+    split_k = config.get("SPLIT_K", 1) if hopper_tuned else 1
+    if split_k > 1:
+        assert split_k & (split_k - 1) == 0
+        partials = torch.empty((split_k, M, N), device=A.device, dtype=torch.float32)
+    else:
+        partials = C
+
     needs_masking = bool(K % config["BLOCK_SIZE_K"] != 0)
 
     def grid(META):
-        return (
+        blocks = triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(
-            triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),
+            N, META["BLOCK_SIZE_N"]
         )
+        return (blocks, split_k) if hopper_tuned else (blocks,)
 
     kernel[grid](
         A,
         B,
-        C,
+        partials,
         As,
         Bs,
         M,
@@ -1617,6 +1760,11 @@ def w8a8_block_fp8_matmul_triton(
         needs_masking=needs_masking,
     )
 
+    if split_k > 1:
+        _reduce_block_fp8_split_k[(triton.cdiv(M * N, 256),)](
+            partials, C, M * N, split_k, 256
+        )
+
     return C