[Example] Add GQA Example

examples/flash_attention/example_gqa_fwd_bshd.py

@@ -0,0 +1,241 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+import torch
+import torch.nn.functional as F
+import tilelang
+from tilelang import Profiler
+from tilelang.autotuner import *
+import tilelang.language as T
+import itertools
+import argparse
+from functools import partial
+
+
+def get_configs():
+    block_M = [128]
+    block_N = [128]
+    num_stages = [2]
+    threads = [256]
+    _configs = list(itertools.product(block_M, block_N, num_stages, threads))
+
+    configs = [{
+        'block_M': c[0],
+        'block_N': c[1],
+        'num_stages': c[2],
+        'threads': c[3]
+    } for c in _configs]
+    return configs
+
+
+def flashattn(batch, heads, seq_len, dim, is_causal, tune=False, groups=1):
+    scale = (1.0 / dim)**0.5 * 1.44269504  # log2(e)
+    head_kv = heads // groups
+    q_shape = [batch, seq_len, heads, dim]
+    kv_shape = [batch, seq_len, head_kv, dim]
+    dtype = "float16"
+    accum_dtype = "float"
+
+    def kernel_func(block_M, block_N, num_stages, threads):
+
+        @T.macro
+        def MMA0(
+            K: T.Buffer(kv_shape, dtype),
+            Q_shared: T.Buffer([block_M, dim], dtype),
+            K_shared: T.Buffer([block_N, dim], dtype),
+            acc_s: T.Buffer([block_M, block_N], accum_dtype),
+            k: T.int32,
+            bx: T.int32,
+            by: T.int32,
+            bz: T.int32,
+        ):
+            T.copy(K[bz, k * block_N:(k + 1) * block_N, by // groups, :], K_shared)
+            if is_causal:
+                for i, j in T.Parallel(block_M, block_N):
+                    acc_s[i, j] = T.if_then_else(bx * block_M + i >= k * block_N + j, 0,
+                                                 -T.infinity(acc_s.dtype))
+            else:
+                T.clear(acc_s)
+            T.gemm(Q_shared, K_shared, acc_s, transpose_B=True, policy=T.GemmWarpPolicy.FullRow)
+
+        @T.macro
+        def MMA1(
+                V: T.Buffer(kv_shape, dtype),
+                V_shared: T.Buffer([block_M, dim], dtype),
+                acc_s_cast: T.Buffer([block_M, block_N], dtype),
+                acc_o: T.Buffer([block_M, dim], accum_dtype),
+                k: T.int32,
+                by: T.int32,
+                bz: T.int32,
+        ):
+            T.copy(V[bz, k * block_N:(k + 1) * block_N, by // groups, :], V_shared)
+            T.gemm(acc_s_cast, V_shared, acc_o, policy=T.GemmWarpPolicy.FullRow)
+
+        @T.macro
+        def Softmax(
+                acc_s: T.Buffer([block_M, block_N], accum_dtype),
+                acc_s_cast: T.Buffer([block_M, block_N], dtype),
+                scores_max: T.Buffer([block_M], accum_dtype),
+                scores_max_prev: T.Buffer([block_M], accum_dtype),
+                scores_scale: T.Buffer([block_M], accum_dtype),
+                scores_sum: T.Buffer([block_M], accum_dtype),
+                logsum: T.Buffer([block_M], accum_dtype),
+        ):
+            T.copy(scores_max, scores_max_prev)
+            T.fill(scores_max, -T.infinity(accum_dtype))
+            T.reduce_max(acc_s, scores_max, dim=1, clear=False)
+            # To do causal softmax, we need to set the scores_max to 0 if it is -inf
+            # This process is called Check_inf in FlashAttention3 code, and it only need to be done
+            # in the first ceil_div(kBlockM, kBlockN) steps.
+            # for i in T.Parallel(block_M):
+            #     scores_max[i] = T.if_then_else(scores_max[i] == -T.infinity(accum_dtype), 0, scores_max[i])
+            for i in T.Parallel(block_M):
+                scores_scale[i] = T.exp2(scores_max_prev[i] * scale - scores_max[i] * scale)
+            for i, j in T.Parallel(block_M, block_N):
+                # Instead of computing exp(x - max), we compute exp2(x * log_2(e) -
+                # max * log_2(e)) This allows the compiler to use the ffma
+                # instruction instead of fadd and fmul separately.
+                acc_s[i, j] = T.exp2(acc_s[i, j] * scale - scores_max[i] * scale)
+            T.reduce_sum(acc_s, scores_sum, dim=1)
+            for i in T.Parallel(block_M):
+                logsum[i] = logsum[i] * scores_scale[i] + scores_sum[i]
+            T.copy(acc_s, acc_s_cast)
+
+        @T.macro
+        def Rescale(
+                acc_o: T.Buffer([block_M, dim], accum_dtype),
+                scores_scale: T.Buffer([block_M], accum_dtype),
+        ):
+            for i, j in T.Parallel(block_M, dim):
+                acc_o[i, j] *= scores_scale[i]
+
+        @T.prim_func
+        def main(
+                Q: T.Buffer(q_shape, dtype),
+                K: T.Buffer(kv_shape, dtype),
+                V: T.Buffer(kv_shape, dtype),
+                Output: T.Buffer(q_shape, dtype),
+        ):
+            with T.Kernel(
+                    T.ceildiv(seq_len, block_M), heads, batch, threads=threads) as (bx, by, bz):
+                Q_shared = T.alloc_shared([block_M, dim], dtype)
+                K_shared = T.alloc_shared([block_N, dim], dtype)
+                V_shared = T.alloc_shared([block_N, dim], dtype)
+                O_shared = T.alloc_shared([block_M, dim], dtype)
+                acc_s = T.alloc_fragment([block_M, block_N], accum_dtype)
+                acc_s_cast = T.alloc_fragment([block_M, block_N], dtype)
+                acc_o = T.alloc_fragment([block_M, dim], accum_dtype)
+                scores_max = T.alloc_fragment([block_M], accum_dtype)
+                scores_max_prev = T.alloc_fragment([block_M], accum_dtype)
+                scores_scale = T.alloc_fragment([block_M], accum_dtype)
+                scores_sum = T.alloc_fragment([block_M], accum_dtype)
+                logsum = T.alloc_fragment([block_M], accum_dtype)
+
+                T.copy(Q[bz, bx * block_M:(bx + 1) * block_M, by, :], Q_shared)
+                T.fill(acc_o, 0)
+                T.fill(logsum, 0)
+                T.fill(scores_max, -T.infinity(accum_dtype))
+
+                loop_range = (
+                    T.min(T.ceildiv(seq_len, block_N), T.ceildiv(
+                        (bx + 1) * block_M, block_N)) if is_causal else T.ceildiv(seq_len, block_N))
+
+                for k in T.Pipelined(loop_range, num_stages=num_stages):
+                    MMA0(K, Q_shared, K_shared, acc_s, k, bx, by, bz)
+                    Softmax(acc_s, acc_s_cast, scores_max, scores_max_prev, scores_scale,
+                            scores_sum, logsum)
+                    Rescale(acc_o, scores_scale)
+                    MMA1(V, V_shared, acc_s_cast, acc_o, k, by, bz)
+                for i, j in T.Parallel(block_M, dim):
+                    acc_o[i, j] /= logsum[i]
+                T.copy(acc_o, O_shared)
+                T.copy(O_shared, Output[bz, bx * block_M:(bx + 1) * block_M, by, :])
+
+        return main
+
+    if tune:
+
+        @autotune(
+            configs=get_configs(),
+            keys=["block_M", "block_N", "num_stages", "threads"],
+            warmup=10,
+            rep=10)
+        @jit(
+            out_idx=[3],
+            supply_type=tilelang.TensorSupplyType.Integer,
+            ref_prog=None,
+            profiler="auto")
+        def kernel(block_M=None, block_N=None, num_stages=None, threads=None):
+            return kernel_func(block_M, block_N, num_stages, threads)
+
+        return kernel()
+    else:
+
+        def kernel(block_M, block_N, num_stages, threads):
+            return kernel_func(block_M, block_N, num_stages, threads)
+
+        return kernel
+
+
+def ref_program(Q, K, V, is_causal, groups=1):
+    # Q: [B, T, HQ, D]
+    # K: [B, T, HK, D]
+    # V: [B, T, HV, D]
+    # HQ = HKV * groups
+    assert Q.size(2) == K.size(
+        2) * groups, f"Q.size(2): {Q.size(2)}, K.size(2): {K.size(2)}, groups: {groups}"
+    assert Q.size(2) == V.size(
+        2) * groups, f"Q.size(2): {Q.size(2)}, V.size(2): {V.size(2)}, groups: {groups}"
+
+    dim = Q.size(-1)
+    K = K.repeat_interleave(groups, dim=2)
+    V = V.repeat_interleave(groups, dim=2)
+    scores = torch.einsum('bqhd,bkhd->bhqk', Q, K)
+    scores = scores / torch.sqrt(torch.tensor(dim, dtype=scores.dtype))
+    if is_causal:
+        seq_len = Q.size(1)
+        mask = torch.tril(torch.ones(seq_len, seq_len, device=scores.device))
+        mask = mask.unsqueeze(0).unsqueeze(0)
+        scores = scores.masked_fill(mask == 0, float('-inf'))
+    attention_weights = F.softmax(scores, dim=-1)
+    output = torch.einsum('bhqk,bkhd->bqhd', attention_weights, V)
+    return output
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=8, help='batch size')
+    parser.add_argument('--heads', type=int, default=32, help='heads')
+    parser.add_argument('--seq_len', type=int, default=4096, help='sequence length')
+    parser.add_argument('--dim', type=int, default=128, help='dim')
+    parser.add_argument('--is_causal', action='store_true', help='causal')
+    parser.add_argument('--tune', action='store_true', help='tune configs')
+    parser.add_argument('--groups', type=int, default=8, help='groups')
+    args = parser.parse_args()
+    batch, heads, seq_len, dim, is_causal, groups = args.batch, args.heads, args.seq_len, args.dim, args.is_causal, args.groups
+    flops_per_matmul = 2.0 * batch * heads * seq_len * seq_len * dim
+    total_flops = 2 * flops_per_matmul
+    if is_causal:
+        total_flops *= 0.5
+
+    if (not args.tune):
+        program = flashattn(
+            batch, heads, seq_len, dim, is_causal, tune=args.tune, groups=groups)(
+                block_M=128, block_N=128, num_stages=1, threads=128)
+        ref_program = partial(ref_program, is_causal=is_causal, groups=groups)
+        mod, params = tilelang.lower(program)
+        mod = Profiler(mod, params, [3], tilelang.TensorSupplyType.Normal)
+        mod.assert_allclose(ref_program, rtol=0.01, atol=0.01)
+        print("All checks pass.")
+        latency = mod.do_bench(ref_program, warmup=500)
+        print("Ref: {:.2f} ms".format(latency))
+        print("Ref: {:.2f} TFlops".format(total_flops / latency * 1e-9))
+        latency = mod.do_bench(mod.func, warmup=500)
+        print("Tile-lang: {:.2f} ms".format(latency))
+        print("Tile-lang: {:.2f} TFlops".format(total_flops / latency * 1e-9))
+    else:
+        best_latency, best_config, _ = flashattn(
+            batch, heads, seq_len, dim, is_causal, tune=args.tune)
+        print(f"Best latency: {best_latency}")
+        print(f"Best TFlops: {total_flops / best_latency * 1e-9}")
+        print(f"Best config: {best_config}")

tilelang/engine/lower.py

@@ -2,7 +2,6 @@
 # Licensed under the MIT License.
 """The compiler for TL programs."""
 
-import tilelang as tl
 import os
 import os.path as osp
 from typing import Union, Optional, Callable
@@ -12,6 +11,10 @@
 from tvm.target import Target
 from tilelang.contrib import hipcc, nvcc
 from tilelang.utils.target import determine_target
+from tilelang.engine.phase import (
+    LowerAndLegalize,
+    OptimizeForTarget,
+)
 
 
 def is_cpu_device_backend(target: Target):
@@ -152,68 +155,12 @@ def lower(
     _is_host_call = get_host_call(is_device_c=is_cpu_device_backend(target))
     _is_device_call = get_device_call(is_device_c=is_cpu_device_backend(target))
 
-    mod = tir.transform.BindTarget(target)(mod)
-
-    mod = tl.transform.FrontendLegalize()(mod)
-    mod = tir.transform.Simplify()(mod)
-    mod = tl.transform.LayoutInference()(mod)
-    mod = tl.transform.LowerTileOp()(mod)
-    mod = tl.transform.LegalizeVectorizedLoop()(mod)
-    mod = tl.transform.LegalizeSafeMemoryAccess()(mod)
-    # Inject Simplify to remove the duplicated conditions
-    mod = tir.transform.Simplify()(mod)
-
-    # which may be introduced by the LegalizeSafeMemoryAccess
-    if target.arch == "sm_90":
-        mod = tl.transform.MultiVersionBuffer()(mod)
-        mod = tl.transform.WarpSpecialized()(mod)
-        mod = tl.transform.InjectSoftwarePipeline()(mod)
-        mod = tir.transform.LowerOpaqueBlock()(mod)
-        # mod = tl.transform.WarpSpecializedPipeline()(mod)
-        mod = tl.transform.InjectFenceProxy()(mod)
-    else:
-        mod = tir.transform.PlanAndUpdateBufferAllocationLocation()(mod)
-        mod = tl.transform.PipelinePlanning()(mod)
-        mod = tl.transform.InjectSoftwarePipeline()(mod)
-
-    mod = tir.transform.LowerOpaqueBlock()(mod)
-    mod = tir.transform.FlattenBuffer()(mod)
-    mod = tir.transform.NarrowDataType(32)(mod)
-    mod = tir.transform.Simplify()(mod)
-    mod = tl.transform.VectorizeLoop()(mod)
-    mod = tir.transform.StorageRewrite()(mod)
-    mod = tir.transform.UnrollLoop()(mod)
-    mod = tir.transform.RenormalizeSplitPattern()(mod)
-    mod = tir.transform.Simplify()(mod)
-    mod = tir.transform.RemoveNoOp()(mod)
-    mod = tir.transform.RewriteUnsafeSelect()(mod)
-    mod = tir.transform.HoistIfThenElse()(mod)
-
-    mod = tir.transform.VerifyMemory()(mod)
-    mod = tir.transform.AnnotateEntryFunc()(mod)
-    # TODO(lei): This is a hack to make sure the
-    # thread level allreduce pass can be applied
-    # in TL. As Tl only use one thread dimension
-    # the var binding information will be lost
-    # in the lowering process with Legalization
-    # and Simplify pass.
-    # We can find a way better to create var instead
-    # of putting the LowerThreadAllreduce before
-    # the Legalization.
-    mod = tl.transform.ThreadPartialSync("shared.dyn")(mod)
-    mod = tir.transform.InferFragment()(mod)
-    mod = tir.transform.LowerThreadAllreduce()(mod)
-    mod = tl.transform.LowerHopperIntrin()(mod)
-    mod = tl.transform.ThreadSync("shared")(mod)
-    mod = tl.transform.ThreadSync("shared.dyn")(mod)
-    mod = tir.transform.InjectPTXAsyncCopy()(mod)
-
-    mod = tl.transform.AnnotateDeviceRegions()(mod)
-    mod = tir.transform.SplitHostDevice()(mod)
-    mod = tir.transform.MergeSharedMemoryAllocations()(mod)
-
-    mod = tl.transform.MakePackedAPI()(mod)
-    mod = tir.transform.LowerDeviceKernelLaunch()(mod)
+    # Phase 1: Lower and legalize the IR
+    mod = LowerAndLegalize(mod, target)
+
+    # Phase 2: Optimize the IR for the target
+    mod = OptimizeForTarget(mod, target)
+
     host_mod = tir.transform.Filter(_is_host_call)(mod)
     host_mod = tir.transform.BindTarget(target_host)(host_mod)
     host_mod = tir.transform.FP8StorageLegalize()(host_mod)