Support Triton FP8 Gemm can handle hidden_dim not divisible by 16 (#9093)

Co-authored-by: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>

python/sglang/srt/layers/quantization/fp8_kernel.py

@@ -1415,3 +1415,221 @@ def per_group_transpose(
         a, trans_a, expert_offsets, k, M_ALIGNMENT, BLOCK_SIZE_M=16, BLOCK_SIZE_K=8
     )
     return trans_a
+
+
+def is_weak_contiguous(x: torch.Tensor):
+    strides = x.stride()
+    sizes = x.shape
+    is_not_transpose = strides[0] == 1 and (strides[1] >= max(1, sizes[0]))
+    is_transpose = strides[1] == 1 and (strides[0] >= max(1, sizes[1]))
+    return is_transpose or is_not_transpose
+
+
+@triton.jit
+def scaled_mm_kernel(
+    a_ptr,
+    b_ptr,
+    scale_a_ptr,
+    scale_b_ptr,
+    c_ptr,
+    bias_ptr,
+    M,
+    N,
+    K,
+    stride_am,
+    stride_ak,
+    stride_bk,
+    stride_bn,
+    stride_cm,
+    stride_cn,
+    ACCUMULATOR_DTYPE: tl.constexpr,
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_N: tl.constexpr,
+    BLOCK_SIZE_K: tl.constexpr,
+    BLOCK_SIZE_SCALE_A: tl.constexpr,
+    BLOCK_SIZE_SCALE_B: tl.constexpr,
+):
+    pid = tl.program_id(axis=0)
+
+    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
+
+    pid_m = pid // num_pid_n
+    pid_n = pid % num_pid_n
+
+    accumulator_dtype = ACCUMULATOR_DTYPE
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=accumulator_dtype)
+
+    # NOTE: Some tensor inputs are so large, they will cause int32 overflow
+    # so it is necessary to use tl.int64 for all the offsets, else SEGV will
+    # eventually occur.
+
+    # Offsets and masks.
+    offsets_am = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
+    masks_am = offsets_am < M
+
+    offsets_bn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)
+    masks_bn = offsets_bn < N
+
+    offsets_k = tl.arange(0, BLOCK_SIZE_K).to(tl.int64)
+    offsets_a = stride_am * offsets_am[:, None] + stride_ak * offsets_k[None, :]
+    offsets_b = stride_bk * offsets_k[:, None] + stride_bn * offsets_bn[None, :]
+
+    # NOTE: BLOCK_SIZE_SCALE_A could be 1 or BLOCK_SIZE_M, so need to create
+    # appropriate offsets and masks for each case. Same goes for
+    # BLOCK_SIZE_SCALE_B.
+    offsets_scale_am = (
+        tl.arange(0, BLOCK_SIZE_SCALE_A)
+        + (BLOCK_SIZE_SCALE_A > 1) * pid_m * BLOCK_SIZE_M
+    )
+    masks_scale_am = offsets_scale_am < M
+
+    offsets_scale_bn = (
+        tl.arange(0, BLOCK_SIZE_SCALE_B)
+        + (BLOCK_SIZE_SCALE_B > 1) * pid_n * BLOCK_SIZE_N
+    )
+    masks_scale_bn = offsets_scale_bn < N
+
+    a_ptrs = a_ptr + offsets_a
+    b_ptrs = b_ptr + offsets_b
+
+    scale_a_ptrs = scale_a_ptr + offsets_scale_am
+    scale_b_ptrs = scale_b_ptr + offsets_scale_bn
+
+    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
+        masks_k = offsets_k < K
+        masks_a = masks_am[:, None] & masks_k[None, :]
+        a = tl.load(a_ptrs, mask=masks_a)
+
+        masks_b = masks_k[:, None] & masks_bn[None, :]
+        b = tl.load(b_ptrs, mask=masks_b)
+
+        # Accumulate results.
+        accumulator = tl.dot(a, b, accumulator, out_dtype=accumulator_dtype)
+
+        offsets_k += BLOCK_SIZE_K
+        a_ptrs += BLOCK_SIZE_K * stride_ak
+        b_ptrs += BLOCK_SIZE_K * stride_bk
+
+    # Apply scale at end.
+    masks_scale_a = masks_scale_am[:, None] & (tl.arange(0, 1) < 1)[:, None]
+    scale_a = tl.load(scale_a_ptrs[:, None], masks_scale_a)
+    # Need to broadcast to the appropriate size, if scale_a is already
+    # (BLOCK_SIZE_M, 1) then it will broadcast to its own shape. Same goes
+    # for scale_b below.
+    scale_a = scale_a.broadcast_to((BLOCK_SIZE_M, 1))
+    accumulator = scale_a * accumulator.to(tl.float32)
+
+    masks_scale_b = masks_scale_bn[:, None] & (tl.arange(0, 1) < 1)[None, :]
+    scale_b = tl.load(scale_b_ptrs[:, None], masks_scale_b)
+    scale_b = scale_b.broadcast_to((BLOCK_SIZE_N, 1))
+    accumulator = scale_b.T * accumulator.to(tl.float32)
+
+    # Convert to output format.
+    c = accumulator.to(c_ptr.type.element_ty)
+
+    # Add bias, it's already in output format, so add it after conversion.
+    if bias_ptr:
+        offsets_bias = offsets_bn
+        bias_ptrs = bias_ptr + offsets_bias
+        bias_mask = offsets_bias < N
+        bias = tl.load(bias_ptrs, bias_mask)
+        c += bias
+
+    # Save output
+    offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)
+    offs_cm = offs_cm.to(tl.int64)
+    offs_cn = offs_cn.to(tl.int64)
+    c_ptrs = c_ptr + 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)
+
+
+# input  - [M, K]
+# weight - [K, N]
+# Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/quantization/compressed_tensors/triton_scaled_mm.py
+def triton_scaled_mm(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    scale_a: torch.Tensor,
+    scale_b: torch.Tensor,
+    out_dtype: type[torch.dtype],
+    bias: Optional[torch.Tensor] = None,
+    block_size_m: int = 32,
+    block_size_n: int = 32,
+    block_size_k: int = 32,
+    use_heuristic=True,
+) -> torch.Tensor:
+    M, K = input.shape
+    N = weight.shape[1]
+
+    assert N > 0 and K > 0 and M > 0
+    assert weight.shape[0] == K
+    assert input.dtype == weight.dtype
+
+    scale_a = scale_a.reshape(-1, 1) if scale_a.dim() <= 1 else scale_a
+    scale_b = scale_b.reshape(-1, 1) if scale_b.dim() <= 1 else scale_b
+
+    assert scale_a.dtype == scale_b.dtype and scale_a.is_floating_point()
+    assert scale_a.shape[1] == 1 and (scale_a.shape[0] == 1 or scale_a.shape[0] == M)
+    assert scale_b.shape[1] == 1 and (scale_b.shape[0] == 1 or scale_b.shape[0] == N)
+    assert out_dtype.is_floating_point
+    assert bias is None or bias.is_floating_point()
+    assert is_weak_contiguous(input)
+    assert is_weak_contiguous(weight)
+
+    grid = lambda META: (
+        triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),
+    )
+
+    result = torch.empty((M, N), dtype=out_dtype, device=input.device)
+
+    has_scalar = lambda x: x.shape[0] == 1 and x.shape[1] == 1
+
+    if use_heuristic:
+        is_small_N = N < 8192
+        next_power_of_2_M = max(32, triton.next_power_of_2(M))
+        if next_power_of_2_M <= 32:
+            tile_shape = (64, 64, 256) if is_small_N else (64, 128, 256)
+        elif next_power_of_2_M <= 64:
+            tile_shape = (64, 64, 256)
+        elif next_power_of_2_M <= 128:
+            tile_shape = (64, 128, 128)
+        else:
+            tile_shape = (128, 128, 128)
+
+    block_size_m, block_size_n, block_size_k = tile_shape
+
+    block_size_sa = 1 if has_scalar(scale_a) else block_size_m
+    block_size_sb = 1 if has_scalar(scale_b) else block_size_n
+
+    accumulator_dtype = tl.float32 if input.is_floating_point() else tl.int32
+
+    # A = input, B = weight, C = result
+    # A = M x K, B = K x N, C = M x N
+    scaled_mm_kernel[grid](
+        input,
+        weight,
+        scale_a,
+        scale_b,
+        result,
+        bias,
+        M,
+        N,
+        K,
+        input.stride(0),
+        input.stride(1),
+        weight.stride(0),
+        weight.stride(1),
+        result.stride(0),
+        result.stride(1),
+        accumulator_dtype,
+        BLOCK_SIZE_M=block_size_m,
+        BLOCK_SIZE_N=block_size_n,
+        BLOCK_SIZE_K=block_size_k,
+        BLOCK_SIZE_SCALE_A=block_size_sa,
+        BLOCK_SIZE_SCALE_B=block_size_sb,
+    )
+
+    return result.to(out_dtype)

python/sglang/srt/layers/quantization/fp8_utils.py

@@ -22,6 +22,7 @@ from sglang.srt.layers.quantization.fp8_kernel import (
     scaled_fp8_quant,
     sglang_per_token_quant_fp8,
     static_quant_fp8,
+    triton_scaled_mm,
     w8a8_block_fp8_matmul_deepgemm,
     w8a8_block_fp8_matmul_triton,
 )
@@ -586,14 +587,25 @@ def apply_fp8_linear(
                 assert (
                     weight_scale.numel() == weight.shape[1]
                 ), "cutlass w8a8 fp8 sgl-kernel only supports per-channel scale"
-                output = fp8_scaled_mm(
-                    qinput,
-                    weight,
-                    x_scale,
-                    weight_scale,
-                    out_dtype=input.dtype,
-                    bias=bias,
+
+                cutlass_compatible_b = (
+                    weight.shape[0] % 16 == 0 and weight.shape[1] % 16 == 0
                 )
+                if not cutlass_compatible_b:
+                    # Massage the input to be 2D
+                    qinput = qinput.view(-1, qinput.shape[-1])
+                    output = triton_scaled_mm(
+                        qinput, weight, x_scale, weight_scale, input.dtype, bias
+                    )
+                else:
+                    output = fp8_scaled_mm(
+                        qinput,
+                        weight,
+                        x_scale,
+                        weight_scale,
+                        out_dtype=input.dtype,
+                        bias=bias,
+                    )
             return output.view(*output_shape)
 
         # torch.scaled_mm supports per tensor weights + activations only

test/srt/quant/test_triton_scaled_mm.py

+import itertools
+import unittest
+from typing import Optional
+
+import torch
+import torch.testing
+
+from sglang.srt.layers.quantization.fp8_kernel import triton_scaled_mm
+from sglang.test.test_utils import CustomTestCase
+
+
+def torch_scaled_mm(
+    a: torch.Tensor,
+    b: torch.Tensor,
+    scale_a: torch.Tensor,
+    scale_b: torch.Tensor,
+    out_dtype: torch.dtype,
+    bias: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Reference implementation using float32 for stability"""
+    out = torch.mm(a.to(torch.float32), b.to(torch.float32))
+    out = scale_a.to(torch.float32) * out * scale_b.to(torch.float32).T
+    if bias is not None:
+        out = out + bias.to(torch.float32)
+    return out.to(out_dtype)
+
+
+class TestScaledMM(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        if not torch.cuda.is_available():
+            raise unittest.SkipTest("This test requires a CUDA device.")
+        torch.set_default_device("cuda")
+
+    def _make_inputs(self, M, K, N, in_dtype):
+        if in_dtype == torch.int8:
+            a = torch.randint(-8, 8, (M, K), dtype=in_dtype, device="cuda")
+            b = torch.randint(-8, 8, (K, N), dtype=in_dtype, device="cuda")
+        else:  # fp8
+            a = torch.clamp(
+                0.1 * torch.randn((M, K), dtype=torch.float16, device="cuda"), -0.3, 0.3
+            ).to(in_dtype)
+            b = torch.clamp(
+                0.1 * torch.randn((K, N), dtype=torch.float16, device="cuda"), -0.3, 0.3
+            ).to(in_dtype)
+        return a, b
+
+    def test_basic_cases(self):
+        """Test core functionality with reduced precision requirements"""
+        test_configs = [
+            (32, 32, 32, torch.int8, torch.float16, False),
+            (64, 64, 64, torch.int8, torch.float16, True),
+        ]
+
+        try:
+            torch.tensor([1.0], dtype=torch.float8_e4m3fn, device="cuda")
+            test_configs.append((32, 32, 32, torch.float8_e4m3fn, torch.float16, False))
+        except:
+            print("FP8 not supported, skipping")
+
+        for M, K, N, in_dtype, out_dtype, with_bias in test_configs:
+            with self.subTest(M=M, K=K, N=N, dtype=in_dtype, bias=with_bias):
+                print(f"Currently testing with in_dtype: {in_dtype}")
+                torch.manual_seed(42)
+
+                input, weight = self._make_inputs(M, K, N, in_dtype)
+                scale_a = 0.1 + 0.05 * torch.rand(
+                    (M, 1), dtype=torch.float32, device="cuda"
+                )
+                scale_b = 0.1 + 0.05 * torch.rand(
+                    (N, 1), dtype=torch.float32, device="cuda"
+                )
+                bias = (
+                    0.01 * torch.randn((M, N), dtype=out_dtype, device="cuda")
+                    if with_bias
+                    else None
+                )
+
+                triton_out = triton_scaled_mm(
+                    input, weight, scale_a, scale_b, out_dtype, bias
+                )
+                ref_out = torch_scaled_mm(
+                    input, weight, scale_a, scale_b, out_dtype, bias
+                )
+
+                # Use relaxed tolerances
+                rtol = 0.15 if in_dtype == torch.int8 else 0.25
+                atol = 0.1 if in_dtype == torch.int8 else 0.15
+
+                torch.testing.assert_close(triton_out, ref_out, rtol=rtol, atol=atol)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)

test/srt/run_suite.py

@@ -57,6 +57,7 @@ suites = {
         TestFile("quant/test_block_int8.py", 22),
         TestFile("quant/test_fp8_kernel.py", 8),
         TestFile("quant/test_int8_kernel.py", 8),
+        TestFile("quant/test_triton_scaled_mm.py", 8),
         TestFile("quant/test_w8a8_quantization.py", 46),
         TestFile("rl/test_update_weights_from_disk.py", 114),
         TestFile("rl/test_update_weights_from_tensor.py", 48),