mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-29 11:16:59 +00:00
[CK] [CK_Tile] Add FMHA scaffolding to CK kernel dispatcher (#5260) ## Motivation The CK Tile dispatcher currently supports GEMM and Grouped Convolution but has no support for Fused Multi-Head Attention (FMHA). The example/ck_tile/01_fmha folder contains a comprehensive FMHA implementation with forward, backward, split-KV, paged-KV, append-KV, and batch-prefill kernels across multiple GPU architectures — but there is no unified dispatch layer for it. This PR ports the FMHA stack into the dispatcher, following the same architectural patterns established by GEMM and Grouped Convolution, enabling runtime kernel selection, JIT compilation from Python, and a declarative C++ example flow. Autotuning heuristics to follow. ## Technical Details This PR adds FMHA scaffolding to the CK dispatcher framework, mirroring GEMM's layered architecture. Seven new C++ runtime headers provide type definitions (coexisting with upstream headers via __has_include, requiring zero modifications to example/ck_tile/01_fmha/), a problem builder with 18+ setters, Signature + Algorithm kernel key matching, a virtual kernel instance, a DECL_FMHA_KERNEL_SET macro with wildcard support and named tile/wave/warp setters, arch-aware registry with JSON export, and a dispatcher with seqtune-aware selection, configurable timing, and multi-stage execution plans for split-KV (two-stage) and backward (three-stage). The codegen pipeline is driven by a fmha_arch_specs.json capturing per-arch tile tables and pipeline constraints for five architectures (gfx90a/942/950/1100/1201), migrated from hardcoded logic in 01_fmha/codegen/, with supporting modules for C++ symbol mappings, validation rules, and named receipt profiles (ck_default, flash, pytorch, aiter, fp32, fp8). Python integration (fmha_utils.py) mirrors the C++ layer with JIT compilation, parallel multi-kernel builds, HIP memory management via ctypes, tolerance-based validation, and a NumPy CPU reference with GQA support. Twenty-seven C++ and thirty-two Python examples cover the full feature surface — forward, split-KV, masks, bias, dropout, GQA, backward, append-KV, batch prefill, fp8, logits soft cap, sink tokens, and parameter sweeps — all JIT-compiled on the fly. ## Test Plan Seven test files cover the runtime types, codegen, and end-to-end correctness. C++ unit tests validate the problem builder, dispatcher planning (single-stage for forward/paged-KV/append-KV; multi-stage for split-KV and backward), registry operations, and the kernel-set declaration macro. Python unit tests verify codegen emission, profile filtering, and 15 validation rules for masks, hdim constraints, and pipeline requirements. GPU execution validation in 01_basic_fmha --validate reports zero errors across 65,536 elements with max absolute error of 7.29e-05. A gold-standard parity suite (test_fmha_parity.py) runs 14 configurations through both the upstream tile_example_fmha_fwd and the dispatcher, comparing exit codes to confirm behavioral parity — all 14 match. ## Test Result The C++ smoke test builds and passes all 9 compiled examples, and a Python JIT sweep (29_sweep_seqlen.py) passes 7/7 configurations reaching up to 375 TFLOPS at seqlen 2048. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Yaswanth Raparti <113389104+yraparti@users.noreply.github.com> Co-authored-by: Mohsen Saffari <mohsen.saffari@amd.com> Co-authored-by: Maksim (Max) Podkorytov <Maksim.Podkorytov@amd.com> Co-authored-by: yashagar <yashagar@amd.com>
268 lines
8.7 KiB
Python
268 lines
8.7 KiB
Python
#!/usr/bin/env python3
|
|
|
|
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
|
|
# SPDX-License-Identifier: MIT
|
|
|
|
"""
|
|
Example 16: Split-KV Attention and Paged KV Cache
|
|
|
|
Demonstrates:
|
|
1. Split-KV: partitioning KV across multiple GPU splits for long sequences
|
|
2. Two-stage execution plan: split (per-partition attention) + combine (merge)
|
|
3. Paged KV cache with configurable page_block_size
|
|
4. CPU reference for split-KV correctness verification
|
|
|
|
Split-KV is critical for long-context inference where seqlen_k >> seqlen_q
|
|
(decoding with long history). Each split processes a chunk of KV independently,
|
|
then partial results are combined with log-sum-exp correction.
|
|
|
|
Usage:
|
|
python3 16_splitkv_fmha.py
|
|
python3 16_splitkv_fmha.py --num-splits 4
|
|
python3 16_splitkv_fmha.py --seqlen-k 2048 --page-size 128
|
|
"""
|
|
|
|
import sys
|
|
import argparse
|
|
from pathlib import Path
|
|
|
|
sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent / "python"))
|
|
import numpy as np
|
|
|
|
from fmha_utils import (
|
|
FmhaProblem,
|
|
FmhaKernelConfig,
|
|
FmhaValidator,
|
|
cpu_attention_fwd,
|
|
detect_gpu_arch,
|
|
setup_fmha_dispatcher,
|
|
)
|
|
|
|
|
|
def cpu_splitkv_attention(
|
|
Q: np.ndarray,
|
|
K: np.ndarray,
|
|
V: np.ndarray,
|
|
scale: float,
|
|
num_splits: int,
|
|
) -> tuple:
|
|
"""CPU reference: split-KV attention with LSE-based combining.
|
|
|
|
Stage 1 (split): Compute partial attention for each KV chunk
|
|
Stage 2 (combine): Merge partial results using log-sum-exp correction
|
|
|
|
Returns: (O_final, partial_Os, partial_lses)
|
|
"""
|
|
batch, nhead, seqlen_q, hdim = Q.shape
|
|
seqlen_k = K.shape[2]
|
|
hdim_v = V.shape[3]
|
|
|
|
chunk_size = (seqlen_k + num_splits - 1) // num_splits
|
|
|
|
partial_Os = np.zeros(
|
|
(num_splits, batch, nhead, seqlen_q, hdim_v), dtype=np.float32
|
|
)
|
|
partial_lses = np.full(
|
|
(num_splits, batch, nhead, seqlen_q), -np.inf, dtype=np.float32
|
|
)
|
|
|
|
for s in range(num_splits):
|
|
k_start = s * chunk_size
|
|
k_end = min(k_start + chunk_size, seqlen_k)
|
|
if k_start >= seqlen_k:
|
|
break
|
|
|
|
K_chunk = K[:, :, k_start:k_end, :]
|
|
V_chunk = V[:, :, k_start:k_end, :]
|
|
|
|
S = np.matmul(Q, K_chunk.transpose(0, 1, 3, 2)) * scale
|
|
S_max = S.max(axis=-1, keepdims=True)
|
|
S_exp = np.exp(S - S_max)
|
|
S_sum = S_exp.sum(axis=-1, keepdims=True)
|
|
|
|
partial_Os[s] = np.matmul(S_exp / S_sum, V_chunk)
|
|
partial_lses[s] = np.log(S_sum.squeeze(-1)) + S_max.squeeze(-1)
|
|
|
|
# Stage 2: Combine using LSE correction
|
|
global_lse = np.max(partial_lses, axis=0) # [batch, nhead, seqlen_q]
|
|
|
|
O_final = np.zeros((batch, nhead, seqlen_q, hdim_v), dtype=np.float32)
|
|
weight_sum = np.zeros((batch, nhead, seqlen_q), dtype=np.float32)
|
|
|
|
for s in range(num_splits):
|
|
correction = np.exp(partial_lses[s] - global_lse)
|
|
correction = correction[..., np.newaxis]
|
|
O_final += partial_Os[s] * correction
|
|
weight_sum += correction.squeeze(-1)
|
|
|
|
O_final = O_final / weight_sum[..., np.newaxis]
|
|
return O_final, partial_Os, partial_lses
|
|
|
|
|
|
def make_page_table(batch: int, seqlen_k: int, page_size: int) -> tuple:
|
|
"""Create a paged KV cache layout.
|
|
|
|
Returns: (page_table, num_pages_per_seq, total_pages)
|
|
"""
|
|
pages_per_seq = (seqlen_k + page_size - 1) // page_size
|
|
total_pages = batch * pages_per_seq
|
|
|
|
page_table = np.arange(total_pages, dtype=np.int32).reshape(batch, pages_per_seq)
|
|
return page_table, pages_per_seq, total_pages
|
|
|
|
|
|
def main():
|
|
parser = argparse.ArgumentParser(description="Split-KV and Paged KV Cache")
|
|
parser.add_argument("--arch", default=detect_gpu_arch())
|
|
parser.add_argument("--batch", type=int, default=2)
|
|
parser.add_argument("--nhead-q", type=int, default=16)
|
|
parser.add_argument("--nhead-k", type=int, default=16)
|
|
parser.add_argument(
|
|
"--seqlen-q", type=int, default=1, help="Typically 1 for decoding"
|
|
)
|
|
parser.add_argument("--seqlen-k", type=int, default=1024)
|
|
parser.add_argument("--hdim", type=int, default=128)
|
|
parser.add_argument("--num-splits", type=int, default=0, help="0 = test multiple")
|
|
parser.add_argument("--page-size", type=int, default=128)
|
|
args = parser.parse_args()
|
|
|
|
print("=" * 70)
|
|
print("Example 16: Split-KV Attention and Paged KV Cache")
|
|
print("=" * 70)
|
|
|
|
sq, sk = args.seqlen_q, args.seqlen_k
|
|
prob = FmhaProblem(
|
|
batch=args.batch,
|
|
nhead_q=args.nhead_q,
|
|
nhead_k=args.nhead_k,
|
|
seqlen_q=sq,
|
|
seqlen_k=sk,
|
|
hdim_q=args.hdim,
|
|
hdim_v=args.hdim,
|
|
)
|
|
|
|
print(
|
|
f"\n Problem: B={prob.batch} Hq={prob.nhead_q} Hk={prob.nhead_k} "
|
|
f"Sq={sq} Sk={sk} D={args.hdim}"
|
|
)
|
|
print(f" Use case: Decoding (Sq={sq} << Sk={sk})")
|
|
|
|
# --- Generate data ---
|
|
np.random.seed(42)
|
|
Q_f32 = (np.random.randn(*prob.q_shape()) * 0.1).astype(np.float32)
|
|
K_f32 = (np.random.randn(*prob.k_shape()) * 0.1).astype(np.float32)
|
|
V_f32 = (np.random.randn(*prob.v_shape()) * 0.1).astype(np.float32)
|
|
Q_fp16 = Q_f32.astype(np.float16)
|
|
K_fp16 = K_f32.astype(np.float16)
|
|
V_fp16 = V_f32.astype(np.float16)
|
|
|
|
# --- Full attention reference ---
|
|
O_full = cpu_attention_fwd(Q_f32, K_f32, V_f32, prob.scale)
|
|
|
|
# --- GPU attempt ---
|
|
print("\n--- GPU Execution ---")
|
|
gpu_output = None
|
|
config = FmhaKernelConfig(
|
|
data_type="fp16",
|
|
hdim_q=args.hdim,
|
|
hdim_v=args.hdim,
|
|
gfx_arch=args.arch,
|
|
)
|
|
setup = setup_fmha_dispatcher(config)
|
|
if not setup.success:
|
|
print(f" JIT build failed: {setup.error}")
|
|
else:
|
|
runner = setup.runner
|
|
print(f" JIT build: {setup.build_time_s:.1f}s")
|
|
res = runner.run(Q_fp16, K_fp16, V_fp16, prob)
|
|
if res.success:
|
|
gpu_output = res.output
|
|
print(f" GPU (full): {res.time_ms:.4f} ms, {res.tflops:.2f} TFLOPS")
|
|
else:
|
|
print(" GPU: Kernel returned failure")
|
|
|
|
# --- Split-KV with various num_splits ---
|
|
print("\n--- Split-KV Execution Plan ---")
|
|
split_configs = [args.num_splits] if args.num_splits > 0 else [1, 2, 3, 4, 8]
|
|
split_configs = [s for s in split_configs if s <= sk]
|
|
|
|
validator = FmhaValidator(rtol=1e-5, atol=1e-5)
|
|
|
|
print("\n Plan stages:")
|
|
print(" Stage 1 (split): Compute partial O and LSE per KV chunk")
|
|
print(" Stage 2 (combine): Merge with exp(lse_i - lse_max) correction")
|
|
|
|
print(
|
|
f"\n {'#':<3} {'Splits':>7} {'ChunkSz':>8} {'Stage1':>8} {'Stage2':>8} "
|
|
f"{'MaxErr':>10} {'Status':>8}"
|
|
)
|
|
print(" " + "-" * 58)
|
|
|
|
for i, ns in enumerate(split_configs, 1):
|
|
chunk_size = (sk + ns - 1) // ns
|
|
|
|
O_split, partial_Os, partial_lses = cpu_splitkv_attention(
|
|
Q_f32,
|
|
K_f32,
|
|
V_f32,
|
|
prob.scale,
|
|
ns,
|
|
)
|
|
|
|
ok, max_abs, _ = validator.check(O_split, O_full)
|
|
tag = "PASS" if ok else "FAIL"
|
|
|
|
print(
|
|
f" {i:<3} {ns:>7} {chunk_size:>8} {'split':>8} {'combine':>8} "
|
|
f"{max_abs:>10.2e} {tag:>8}"
|
|
)
|
|
|
|
# --- Paged KV Cache ---
|
|
print("\n--- Paged KV Cache ---")
|
|
page_sizes = [64, 128, 256]
|
|
|
|
print(
|
|
f"\n {'PageSize':>9} {'Pages/Seq':>10} {'TotalPages':>11} {'Utilization':>12}"
|
|
)
|
|
print(" " + "-" * 46)
|
|
|
|
for ps in page_sizes:
|
|
pt, pps, tp = make_page_table(args.batch, sk, ps)
|
|
used_slots = args.batch * sk
|
|
total_slots = tp * ps
|
|
util = used_slots / total_slots * 100
|
|
print(f" {ps:>9} {pps:>10} {tp:>11} {util:>11.1f}%")
|
|
|
|
print(f"\n Page table example (batch=0, page_size={args.page_size}):")
|
|
pt, pps, _ = make_page_table(args.batch, sk, args.page_size)
|
|
pages_str = ", ".join(str(p) for p in pt[0, : min(8, pps)])
|
|
if pps > 8:
|
|
pages_str += f" ... ({pps} pages)"
|
|
print(f" [{pages_str}]")
|
|
print(" Maps logical KV positions -> physical page indices")
|
|
|
|
# --- GPU validation if available ---
|
|
if gpu_output is not None:
|
|
print("\n--- GPU vs Full-Attention Reference ---")
|
|
val = FmhaValidator(rtol=1e-2, atol=1e-2)
|
|
ok, max_abs, max_rel = val.check(gpu_output, O_full)
|
|
print(
|
|
f" max_abs={max_abs:.2e}, max_rel={max_rel:.2e}, {'PASS' if ok else 'FAIL'}"
|
|
)
|
|
|
|
# --- Summary ---
|
|
print("\n" + "=" * 70)
|
|
print(f" Split-KV: Partitions seqlen_k={sk} across splits")
|
|
print(" Plan: 2-stage (split partial O/LSE -> combine with correction)")
|
|
print(f" Paged KV: page_block_size={args.page_size} ({pps} pages/seq)")
|
|
print(" Use case: Long-context decoding (Sq << Sk)")
|
|
print(" GPU: Prebuilt kernel runs full attention (no split-KV)")
|
|
print(" Status: PASS (CPU split-KV matches full attention)")
|
|
print("=" * 70)
|
|
|
|
return 0
|
|
|
|
|
|
if __name__ == "__main__":
|
|
sys.exit(main())
|