Coverage for src/flag_gems/ops/w8a8_block_fp8_matmul.py: 49%
82 statements
« prev ^ index » next coverage.py v7.6.9, created at 2026-03-25 02:48 +0800
« prev ^ index » next coverage.py v7.6.9, created at 2026-03-25 02:48 +0800
1import functools
2import json
3import logging
4import os
5from typing import Any, Dict, List, Optional
7import torch
8import triton
9import triton.language as tl
11logger = logging.getLogger(__name__)
14@triton.jit
15def w8a8_block_fp8_matmul_kernel(
16 A,
17 B,
18 C,
19 As,
20 Bs,
21 M,
22 N,
23 K,
24 group_n,
25 group_k,
26 stride_am,
27 stride_ak,
28 stride_bk,
29 stride_bn,
30 stride_cm,
31 stride_cn,
32 stride_As_m,
33 stride_As_k,
34 stride_Bs_k,
35 stride_Bs_n,
36 BLOCK_SIZE_M: tl.constexpr,
37 BLOCK_SIZE_N: tl.constexpr,
38 BLOCK_SIZE_K: tl.constexpr,
39 GROUP_SIZE_M: tl.constexpr,
40):
41 pid = tl.program_id(axis=0)
42 num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
43 num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
44 num_pid_in_group = GROUP_SIZE_M * num_pid_n
45 group_id = pid // num_pid_in_group
46 first_pid_m = group_id * GROUP_SIZE_M
47 group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
48 pid_m = first_pid_m + (pid % group_size_m)
49 pid_n = (pid % num_pid_in_group) // group_size_m
51 offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
52 offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
53 offs_k = tl.arange(0, BLOCK_SIZE_K)
54 a_ptrs = A + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak)
55 b_ptrs = B + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)
57 As_ptrs = As + offs_am * stride_As_m
58 offs_bsn = offs_bn // group_n
59 Bs_ptrs = Bs + offs_bsn * stride_Bs_n
61 accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
62 for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
63 a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0)
64 b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)
66 k_start = k * BLOCK_SIZE_K
67 offs_ks = k_start // group_k
68 a_s = tl.load(As_ptrs + offs_ks * stride_As_k)
69 b_s = tl.load(Bs_ptrs + offs_ks * stride_Bs_k)
71 accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :]
72 a_ptrs += BLOCK_SIZE_K * stride_ak
73 b_ptrs += BLOCK_SIZE_K * stride_bk
75 if C.dtype.element_ty == tl.bfloat16:
76 c = accumulator.to(tl.bfloat16)
77 elif C.dtype.element_ty == tl.float16:
78 c = accumulator.to(tl.float16)
79 else:
80 c = accumulator.to(tl.float32)
82 offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
83 offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
84 c_ptrs = C + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
85 c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
86 tl.store(c_ptrs, c, mask=c_mask)
89@functools.lru_cache
90def get_w8a8_block_fp8_configs(
91 N: int, K: int, block_n: int, block_k: int
92) -> Optional[Dict[int, Any]]:
93 device_name = torch.cuda.get_device_name().replace(" ", "_")
94 json_file_name = (
95 f"N={N},K={K},device_name={device_name},"
96 f"dtype=fp8_w8a8,block_shape=[{block_n},{block_k}].json"
97 )
99 config_dir = os.path.join(os.path.dirname(__file__), "..", "utils", "configs")
100 config_file_path = os.path.join(config_dir, json_file_name)
102 if os.path.exists(config_file_path):
103 with open(config_file_path) as f:
104 logger.info(
105 "Using configuration from %s for W8A8 Block FP8 kernel.",
106 config_file_path,
107 )
108 return {int(key): val for key, val in json.load(f).items()}
110 logger.warning(
111 "Using default W8A8 Block FP8 kernel config. Performance might "
112 "be sub-optimal! Config file not found at %s",
113 config_file_path,
114 )
115 return None
118def w8a8_block_fp8_matmul(
119 A: torch.Tensor,
120 B: torch.Tensor,
121 As: torch.Tensor,
122 Bs: torch.Tensor,
123 block_size: List[int],
124 output_dtype: torch.dtype = torch.float16,
125) -> torch.Tensor:
126 assert len(block_size) == 2
127 block_n, block_k = block_size[0], block_size[1]
129 assert A.shape[-1] == B.shape[-1]
130 assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous()
131 assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
132 M = A.numel() // A.shape[-1]
134 assert B.ndim == 2 and Bs.ndim == 2
135 N, K = B.shape
136 assert triton.cdiv(N, block_n) == Bs.shape[0]
137 assert triton.cdiv(K, block_k) == Bs.shape[1]
139 C_shape = A.shape[:-1] + (N,)
140 C = A.new_empty(C_shape, dtype=output_dtype)
142 configs = get_w8a8_block_fp8_configs(N, K, block_n, block_k)
143 if configs:
144 config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
145 else:
146 config = {
147 "BLOCK_SIZE_M": 64,
148 "BLOCK_SIZE_N": block_size[0],
149 "BLOCK_SIZE_K": block_size[1],
150 "GROUP_SIZE_M": 32,
151 "num_warps": 4,
152 "num_stages": 2,
153 }
155 def grid(META):
156 return (
157 triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),
158 )
160 w8a8_block_fp8_matmul_kernel[grid](
161 A,
162 B,
163 C,
164 As,
165 Bs,
166 M,
167 N,
168 K,
169 block_n,
170 block_k,
171 A.stride(-2),
172 A.stride(-1),
173 B.stride(1),
174 B.stride(0),
175 C.stride(-2),
176 C.stride(-1),
177 As.stride(-2),
178 As.stride(-1),
179 Bs.stride(1),
180 Bs.stride(0),
181 **config,
182 )
184 return C