Update Cuda Blackwell

benchmarks/benchmark_flash_attention_fa3.py

@@ -0,0 +1,117 @@
+# Install the newest triton version with
+# pip install "git+https://github.com/openai/triton.git#egg=triton&subdirectory=python"
+import pickle
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from einops import rearrange, repeat
+
+from flash_attn.utils.benchmark import benchmark_all, benchmark_forward, benchmark_backward
+from flash_attn.utils.benchmark import benchmark_fwd_bwd, benchmark_combined
+
+from flash_attn import flash_attn_func
+
+try:
+    from triton.ops.flash_attention import attention as attention_triton
+except ImportError:
+    attention_triton = None
+
+try:
+    import xformers.ops as xops
+except ImportError:
+    xops = None
+
+
+def flops(batch, seqlen, headdim, nheads, causal, mode="fwd"):
+    assert mode in ["fwd", "bwd", "fwd_bwd"]
+    f = 4 * batch * seqlen**2 * nheads * headdim // (2 if causal else 1)
+    return f if mode == "fwd" else (2.5 * f if mode == "bwd" else 3.5 * f)
+
+
+def efficiency(flop, time):
+    return (flop / time / 10**12) if not math.isnan(time) else 0.0
+
+
+def attention_pytorch(qkv, dropout_p=0.0, causal=True):
+    """
+    Arguments:
+        qkv: (batch_size, seqlen, 3, nheads, head_dim)
+        dropout_p: float
+    Output:
+        output: (batch_size, seqlen, nheads, head_dim)
+    """
+    batch_size, seqlen, _, nheads, d = qkv.shape
+    q, k, v = qkv.unbind(dim=2)
+    q = rearrange(q, 'b t h d -> (b h) t d')
+    k = rearrange(k, 'b s h d -> (b h) d s')
+    softmax_scale = 1.0 / math.sqrt(d)
+    # Preallocate attn_weights for `baddbmm`
+    scores = torch.empty(batch_size * nheads, seqlen, seqlen,
+                         dtype=qkv.dtype, device=qkv.device)
+    scores = rearrange(torch.baddbmm(scores, q, k, beta=0, alpha=softmax_scale),
+                       '(b h) t s -> b h t s', h=nheads)
+    if causal:
+        # "triu_tril_cuda_template" not implemented for 'BFloat16'
+        # So we have to construct the mask in float
+        causal_mask = torch.triu(torch.full(
+            (seqlen, seqlen), -10000.0, device=scores.device), 1)
+        # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+        scores = scores + causal_mask.to(dtype=scores.dtype)
+    attention = torch.softmax(scores, dim=-1)
+    attention_drop = F.dropout(attention, dropout_p)
+    output = torch.einsum('bhts,bshd->bthd', attention_drop, v)
+    return output.to(dtype=qkv.dtype)
+
+
+def time_fwd_bwd(func, *args, **kwargs):
+    time_f, time_b = benchmark_fwd_bwd(func, *args, **kwargs)
+    return time_f[1].mean, time_b[1].mean
+
+
+repeats = 30
+device = 'cuda'
+dtype = torch.bfloat16
+
+bs_seqlen_vals = [(32, 512), (16, 1024), (8, 2048), (4, 4096), (2, 8192), (1, 16384)]
+causal_vals = [False, True]
+headdim_vals = [128]
+nheads = 16
+dropout_p = 0.0
+
+methods = (["Flash"])
+
+time_f = {}
+time_b = {}
+time_f_b = {}
+speed_f = {}
+speed_b = {}
+speed_f_b = {}
+for causal in causal_vals:
+    for headdim in headdim_vals:
+        for batch_size, seqlen in bs_seqlen_vals:
+            config = (causal, headdim, batch_size, seqlen)
+            q = torch.randn(batch_size, seqlen, nheads, headdim,
+                            device=device, dtype=dtype, requires_grad=True)
+            k = torch.randn(batch_size, seqlen, nheads, headdim,
+                            device=device, dtype=dtype, requires_grad=True)
+            v = torch.randn(batch_size, seqlen, nheads, headdim,
+                            device=device, dtype=dtype, requires_grad=True)
+
+            f, b = time_fwd_bwd(
+                flash_attn_func, q, k, v, dropout_p, causal=causal, repeats=repeats, verbose=False
+            )
+            time_f[config, "Flash"] = f
+            time_b[config, "Flash"] = b
+
+            print(
+                f"[b, s, h, d] = [{batch_size}, {seqlen}, {nheads}, {headdim}], causal={causal}")
+            for method in methods:
+                speed_b[config, method] = efficiency(
+                    flops(batch_size, seqlen, headdim,
+                          nheads, causal, mode="bwd"),
+                    time_b[config, method]
+                )
+                print(f"bwd: {speed_b[config, method]:.2f} TFLOPs/s")
+

csrc/flash_attn_ck/mha_bwd.cpp

@@ -23,7 +23,10 @@ fmha_bwd_traits get_ck_fmha_bwd_traits(const mask_info &mask,
                            false,    // has_dbias
                            has_dropout,
                            false, // s_randval
-                           deterministic};
+                           deterministic,
+                           true, // uses_ext_asm
+                           true, // is_v3_atomic_fp32
+                           1}; // how_v3_bf16_cvt 0:RTNE; 1:RTNA; 2:RTZ
 }
 
 fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
@@ -99,11 +102,11 @@ fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
     ck_tile::index_t stride_dv = dv.stride(1);
     ck_tile::index_t nhead_stride_dv = dv.stride(2);
 
-    // dq_acc: (split, batch_size, seqlen_q, nheads, hdim)
+    // dq_acc: (split, batch_size, nheads, seqlen_q, hdim)
     ck_tile::index_t split_stride_dq_acc = dq_acc.stride(0);
     ck_tile::index_t batch_stride_dq_acc = dq_acc.stride(1);
-    ck_tile::index_t stride_dq_acc = dq_acc.stride(2);
-    ck_tile::index_t nhead_stride_dq_acc = dq_acc.stride(3);
+    ck_tile::index_t nhead_stride_dq_acc = dq_acc.stride(2);
+    ck_tile::index_t stride_dq_acc = dq_acc.stride(3);
 
     float p_undrop = 1.0 - p_dropout;
 
@@ -191,7 +194,7 @@ fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
                          static_cast<ck_tile::index_t>(mask.type),
                          p_dropout,
                          p_undrop,
-                         {drop_seed, drop_offset}};
+                         std::make_pair(drop_seed, drop_offset)};
 }
 
 std::vector<at::Tensor>
@@ -318,11 +321,11 @@ mha_bwd(const at::Tensor &dout,                   // batch_size x seqlen_q x num
     at::Tensor dq_accum;
 
     if (!deterministic) {
-        dq_accum = torch::zeros({1, batch_size, seqlen_q, num_heads, head_size}, opts.dtype(at::kFloat));
+        dq_accum = torch::zeros({1, batch_size, num_heads, seqlen_q, head_size}, opts.dtype(at::kFloat));
     } else {
         const ck_tile::index_t kN0 = head_size <= 128 ? 128 : 64;
         const ck_tile::index_t nsplits = ck_tile::integer_divide_ceil(seqlen_k, kN0);
-        dq_accum = torch::zeros({nsplits, batch_size, seqlen_q, num_heads, head_size}, opts.dtype(at::kFloat));
+        dq_accum = torch::zeros({nsplits, batch_size, num_heads, seqlen_q, head_size}, opts.dtype(at::kFloat));
     }
 
     at::Tensor dk_expanded, dv_expanded;
@@ -399,4 +402,4 @@ mha_bwd(const at::Tensor &dout,                   // batch_size x seqlen_q x num
     }
 
     return { dq, dk, dv, softmax_d };
-}
+}

csrc/flash_attn_ck/mha_fwd.cpp

@@ -137,7 +137,7 @@ fmha_fwd_args get_ck_fmha_fwd_args(bool has_lse,
                          static_cast<ck_tile::index_t>(mask.type),
                          p_dropout,
                          has_dropout_randval,
-                         {drop_seed, drop_offset}};
+                         std::make_pair(drop_seed, drop_offset)};
 }
 
 std::vector<at::Tensor>

csrc/flash_attn_ck/mha_varlen_bwd.cpp

@@ -23,7 +23,10 @@ fmha_bwd_traits get_ck_fmha_varlen_bwd_traits(const mask_info &mask,
                            false,    // has_dbias
                            has_dropout,
                            false, // s_randval
-                           deterministic};
+                           deterministic,
+                           false, // uses_ext_asm
+                           head_size != 64, // is_v3_atomic_fp32
+                           2}; // how_v3_bf16_cvt 0:RTNE; 1:RTNA; 2:RTZ
 }
 
 fmha_bwd_args get_ck_fmha_varlen_bwd_args(const mask_info &mask,
@@ -197,7 +200,7 @@ fmha_bwd_args get_ck_fmha_varlen_bwd_args(const mask_info &mask,
                          static_cast<ck_tile::index_t>(mask.type),
                          p_dropout,
                          p_undrop,
-                         {drop_seed, drop_offset}};
+                         std::make_pair(drop_seed, drop_offset)};
 }
 
 std::vector<at::Tensor>
@@ -426,4 +429,4 @@ mha_varlen_bwd(const at::Tensor &dout,                   // total_q x num_heads
     }
 
     return { dq, dk, dv, softmax_d };
-}
+}

csrc/flash_attn_ck/mha_varlen_fwd.cpp

@@ -140,7 +140,7 @@ fmha_fwd_args get_ck_fmha_varlen_fwd_args(bool has_lse,
                          static_cast<ck_tile::index_t>(mask.type),
                          p_dropout,
                          has_dropout_randval,
-                         {drop_seed, drop_offset}};
+                         std::make_pair(drop_seed, drop_offset)};
 }
 
 std::vector<at::Tensor>

flash_attn/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "2.6.3"
+__version__ = "3.0.0.r1"
 
 from flash_attn.flash_attn_interface import (
     flash_attn_func,

setup.py

@@ -348,6 +348,8 @@ def validate_and_update_archs(archs):
         f"build/fmha_*wd*.cpp"
     )
 
+    sources+=glob.glob(f"csrc/composable_kernel/example/ck_tile/01_fmha/hsaco/*.cpp")
+
     rename_cpp_to_cu(sources)
 
     renamed_sources = ["csrc/flash_attn_ck/flash_api.cu",
@@ -358,6 +360,8 @@ def validate_and_update_archs(archs):
                        "csrc/flash_attn_ck/mha_varlen_bwd.cu",
                        "csrc/flash_attn_ck/mha_varlen_fwd.cu"] + glob.glob(f"build/fmha_*wd*.cu")
 
+    renamed_sources+=glob.glob(f"csrc/composable_kernel/example/ck_tile/01_fmha/hsaco/*.cu")
+
     cc_flag += ["-O3","-std=c++17",
                 "-DCK_TILE_FMHA_FWD_FAST_EXP2=1",
                 "-fgpu-flush-denormals-to-zero",