Add more flex attention cases to benchmark. (#3928)

chengjunlu · web-flow · commit 94b34733c0fc · 2025-04-28T10:03:08.000+08:00
Add the flex attention shapes which is used by real model to benchmark
for tracking performance.

I commented out 4 cases for now for the reason:
1. There is not enough share local memory for the Triton kernel. 
- Append shapes of Deepseek-v3 (Nope)
- Decode shapes of Deepseek-v3 (Nope)
2. Flex Attention doesn't support such kind of shapes: Error:
LoweringException: ValueError: Number of shared query heads sharing the
same KV head must be power of 2.
- Decode shapes of Qwen2-7B
3. Triton kernel block shapes must be power of 2.
- Decode shapes of Phi3-mini-3.8B

We will investigate the first issue on Triton side later.

Signed-off-by: Lu,Chengjun &lt;chengjun.lu@intel.com&gt;
diff --git a/.github/workflows/triton-benchmarks.yml b/.github/workflows/triton-benchmarks.yml
@@ -281,8 +281,7 @@ jobs:
           python flex_attention_benchmark_causal_mask.py --reports $REPORTS --n_runs $N_RUNS
 
           source ../../scripts/capture-hw-details.sh
-          python build_report.py $REPORTS/flexAttnCausal-performance.csv $REPORTS/flexAttnCausal-triton-report.csv --benchmark flexAttnCausal --compiler triton --param_cols "Z,H,N_CTX,D_HEAD,CAUSAL" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
-          python build_report.py $REPORTS/flexAttnCausal-performance.csv $REPORTS/flexAttnCausal-xetla-report.csv --benchmark flexAttnCausal --compiler xetla --param_cols "Z,H,N_CTX,D_HEAD,CAUSAL" --tflops_col XeTLA-TFlops --hbm_col "XeTLA-GB/s" --tag $TAG
+          python build_report.py $REPORTS/flexAttnCausal-performance.csv $REPORTS/flexAttnCausal-triton-report.csv --benchmark flexAttnCausal --compiler triton --param_cols "Z,H_q,H_kv,N_CTX_q,N_CTX_kv,D_HEAD_qk,D_HEAD_v" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
 
       - name: Run Triton FlexAttention Custom Masks fwd kernel benchmark
         if: ${{ steps.install.outcome == 'success' && !cancelled() && (inputs.benchmarks == '' || contains(fromJson(inputs.benchmarks || '[]'), 'flex_attention_benchmark_custom_masks.py')) && !contains(fromJson(inputs.skip_benchmarks || '[]'), 'flex_attention_benchmark_custom_masks.py') }}
diff --git a/benchmarks/triton_kernels_benchmark/flex_attention_benchmark_causal_mask.py b/benchmarks/triton_kernels_benchmark/flex_attention_benchmark_causal_mask.py
@@ -3,18 +3,19 @@
 import os
 from torch.nn.attention.flex_attention import (
     create_block_mask,
+    create_mask,
     flex_attention,
 )
 
 import torch
 import torch.nn.functional as F
+
 import triton_kernels_benchmark as benchmark_suit
-from triton_kernels_benchmark import xetla_kernel
 
 torch._dynamo.config.recompile_limit = 100  # pylint: disable=protected-access
 
 # Compile the flex_attention function
-flex_attention = torch.compile(flex_attention, dynamic=False)
+compiled_flex_attention = torch.compile(flex_attention, dynamic=False)
 
 
 @lru_cache
@@ -27,112 +28,127 @@ def causal_mask(_, __, q_idx, kv_idx):
     return q_idx >= kv_idx
 
 
+throughput_test = os.getenv('THROUGHPUT_TEST', '0') == '1'
+batch_sizes = [16, 32, 64] if throughput_test else [1]
+
+
 # Kernel profiling for Backward mode is not working as expected:
 # For details: https://github.com/pytorch/pytorch/issues/144778
 @benchmark_suit.perf_report(
     benchmark_suit.Benchmark(
-        x_names=['Z', 'H', 'N_CTX', 'D_HEAD', 'CAUSAL', 'MODE'],
-        x_vals=[[z, h, 16384 // z, dhead, causal, mode]
-                for z in [1, 2, 4, 8, 16, 32]
-                for (h, dhead) in [(16, 128), (32, 64)]
-                for causal in [True]
-                for mode in [os.getenv('FA_KERNEL_MODE', 'fwd')]]  #
-        + [[4, 48, 1024, 64, True, mode] for mode in [os.getenv('FA_KERNEL_MODE', 'fwd')]]  #
-        + [[z, h, 1024, dhead, True, mode]
-           for z in [1, 2, 4, 8, 16, 32, 64]
-           for (h, dhead) in [(8, 128), (32, 96), (4, 128)]
-           for mode in [os.getenv('FA_KERNEL_MODE', 'fwd')]],
+        x_names=['Z', 'H_q', 'H_kv', 'N_CTX_q', 'N_CTX_kv', 'D_HEAD_qk', 'D_HEAD_v', 'MODE'],
+        x_vals=
+        # Multi-head attention. H_q equals H_kv
+        # Prefill shapes of Phi3-mini-3.8B
+        [[z, 32, 32, 1024, 1024, 96, 96, 'fwd'] for z in batch_sizes] +
+        # Prefill shapes of Deepseek-v3
+        [[z, 128, 128, 1024, 1024, 192, 128, 'fwd'] for z in batch_sizes] +
+        # Append shapes of Phi3-mini-3.8B
+        [[z, 32, 32, 512, 1024 + 128 + 512, 96, 96, 'fwd'] for z in batch_sizes] +
+
+        # Multi-query attention. H_kv equals 1.
+        # Append shapes of Deepseek-v3 (Nope)
+        [
+            # RuntimeError: No valid triton configs. OutOfResources: out of resource: shared memory, Required: 133120, Hardware limit: 131072.
+            # [z, 128, 1, 512, 1024 + 128 + 512, 64, 512, 'fwd'] for z in batch_sizes
+        ] +
+        # Append shapes of Deepseek-v3 (Rope)
+        [] +
+
+        # Grouped-query attention. H_q / H_kv > 1
+        # Prefill shapes of Llama-3.1-8B
+        [[z, 32, 8, 1024, 1024, 128, 128, 'fwd'] for z in batch_sizes] +
+        # Prefill shapes of Qwen2-7B
+        [[z, 28, 4, 1024, 1024, 128, 128, 'fwd'] for z in batch_sizes] +
+        # Append shapes of Llama-3.1-8B
+        [[z, 32, 8, 512, 1024 + 128 + 512, 128, 128, 'fwd'] for z in batch_sizes] +
+        # Append shapes of Qwen2-7B
+        [[z, 28, 4, 512, 1024 + 128 + 512, 128, 128, 'fwd'] for z in batch_sizes] +
+
+        # FlexDecoding configuration. N_CTX_q equals 1. N_CTX_kv >= 1k
+        # Decode shapes of Llama-3.1-8B
+        [[z, 32, 8, 1, 1024 + 64, 128, 128, 'fwd'] for z in batch_sizes] +
+        # Decode shapes of Phi3-mini-3.8B
+        [
+            # acc = acc.reshape(G, BLOCK_M_PER_HQ, V_HEAD_DIM)
+            # ValueError: Shape element 2 must be a power of 2
+            # [z, 32, 32, 1, 1024 + 64, 96, 96, 'fwd'] for z in batch_sizes
+        ] +
+        # Decode shapes of Qwen2-7B
+        [
+            # torch._inductor.exc.InductorError: LoweringException: ValueError: Number of shared query heads sharing the same KV head must be power of 2.
+            # [z, 28, 4, 1, 1024 + 64, 128, 128, 'fwd'] for z in batch_sizes
+        ] +
+        # Decode shapes of Deepseek-v3 (Nope)
+        [
+            # RuntimeError: No valid triton configs. OutOfResources: out of resource: shared memory, Required: 264192, Hardware limit: 131072.
+            # [z, 128, 1, 1, 1024, 64, 512, 'fwd'] for z in batch_sizes
+        ] +
+        # Decode shapes of Deepseek-v3 (Rope)
+        [],
         line_arg='provider',
-        line_vals=['triton', 'xetla'],
-        line_names=['Triton', 'XeTLA'],
+        line_vals=['triton'],
+        line_names=['Triton'],
         styles=[('green', '-'), ('green', '--'), ('blue', '-'), ('blue', '--')],
         ylabel=['GB/s', 'TFlops'],
         plot_name='flexAttnCausal-performance',
         args={},
     ))
-def benchmark(Z, H, N_CTX, D_HEAD, CAUSAL, MODE, provider):
-    assert MODE in ['fwd', 'bwd']
-    assert CAUSAL
+def benchmark(Z, H_q, H_kv, N_CTX_q, N_CTX_kv, D_HEAD_qk, D_HEAD_v, MODE, provider):
+    assert MODE in ['fwd']
     dtype = torch.float16
-    q = torch.randn((Z, H, N_CTX, D_HEAD), device='xpu', dtype=dtype, requires_grad=True)
-    k = torch.randn((Z, H, N_CTX, D_HEAD), device='xpu', dtype=dtype, requires_grad=True)
-    v = torch.randn((Z, H, N_CTX, D_HEAD), device='xpu', dtype=dtype, requires_grad=True)
+    q = torch.randn((Z, H_q, N_CTX_q, D_HEAD_qk), device='xpu', dtype=dtype, requires_grad=MODE == 'bwd')
+    k = torch.randn((Z, H_kv, N_CTX_kv, D_HEAD_qk), device='xpu', dtype=dtype, requires_grad=MODE == 'bwd')
+    v = torch.randn((Z, H_kv, N_CTX_kv, D_HEAD_v), device='xpu', dtype=dtype, requires_grad=MODE == 'bwd')
     sm_scale = 0.125
     if MODE == 'bwd':
         sm_scale = 1.3
 
     quantiles = [0.5, 0.0, 1.0]
     if provider == 'triton':
-        kernel_options = {'num_stages': 2, 'num_warps': 16 if D_HEAD == 128 else 8, 'BLOCKS_ARE_CONTIGUOUS': True}
-        block_mask = create_block_mask_cached(causal_mask, 1, 1, N_CTX, N_CTX, device=q.device)
-        triton_fn = lambda: flex_attention(q, k, v, block_mask=block_mask, scale=sm_scale, kernel_options=kernel_options
-                                           )
+        kernel_options = {'num_stages': 2, 'num_warps': 16 if D_HEAD_qk == 128 else 8, 'BLOCKS_ARE_CONTIGUOUS': True}
+        block_mask = create_block_mask_cached(causal_mask, 1, 1, N_CTX_q, N_CTX_kv, device='xpu')
+        triton_fn = lambda: compiled_flex_attention(q, k, v, block_mask=block_mask, scale=sm_scale, enable_gqa=(
+            not H_q == H_kv), kernel_options=kernel_options)
+        torch_fn = lambda: flex_attention(q, k, v, block_mask=block_mask, scale=sm_scale, enable_gqa=not H_q == H_kv)
         if MODE == 'bwd':
             triton_o = triton_fn()
             triton_do = torch.randn_like(triton_o)
             triton_fn = lambda: triton_o.backward(triton_do, retain_graph=True)
-        torch_fn = lambda: F.scaled_dot_product_attention(q.cpu(), k.cpu(), v.cpu(), is_causal=True, scale=sm_scale).to(
-            torch.float32)
-        if MODE == 'bwd':
-            torch_o = torch_fn()
-            torch_do = torch.randn_like(torch_o)
-            torch_fn = lambda: torch_o.backward(torch_do, retain_graph=True)
-        if MODE == 'fwd':
-            atol = 1e-1 if N_CTX == 16384 else 1e-2
-            benchmark_suit.assert_close(triton_fn, torch_fn, atol=atol, rtol=1e-3, err_msg='triton to torch')
-        else:
-            benchmark_suit.assert_close(lambda: triton_o, lambda: torch_o, atol=1e-2, rtol=0, err_msg='triton to torch')
+
+        benchmark_suit.assert_close(triton_fn, torch_fn, atol=1e-2, rtol=1e-3, err_msg='triton to torch')
         _, min_ms, max_ms, mean, cv = benchmark_suit.do_bench(triton_fn, n_warmup=10, n_repeat=10, quantiles=quantiles)
 
-    elif provider == 'xetla':
-        xetla_fn = None
-        if MODE == 'fwd':
-            module_name = 'flash_attn_causal_True'.lower()
-            func = getattr(xetla_kernel, module_name)
-            out = torch.empty_like(q, device='xpu', dtype=dtype)
-            size_score = Z * H * N_CTX * N_CTX
-            size_attn_mask = Z * N_CTX * N_CTX
-            dropout_mask = torch.empty((size_score, ), device='xpu', dtype=torch.uint8)
-            bias = torch.empty((size_attn_mask, ), device='xpu', dtype=dtype)
-            size_ml = Z * H * N_CTX
-            m = torch.empty((size_ml, ), device='xpu', dtype=torch.float)
-            l = torch.empty((size_ml, ), device='xpu', dtype=torch.float)
-            xetla_fn = lambda: func(q, k, v, out, dropout_mask, bias, m, l, Z, H, D_HEAD, N_CTX, N_CTX, sm_scale)
-        if MODE == 'bwd':
-            module_name = 'flash_attn_bwd_causal_True'.lower()
-            func = getattr(xetla_kernel, module_name)
-            grad_out = torch.empty_like(q, device='xpu', dtype=dtype, requires_grad=True)
-            bias = torch.empty_like(q, device='xpu', dtype=dtype, requires_grad=True)
-            dropout = torch.empty_like(q, device='xpu', dtype=torch.uint8)
-            out = torch.empty_like(q, device='xpu', dtype=dtype, requires_grad=True)
-            log_sumexp = torch.zeros(q.size(), device='xpu', dtype=dtype, requires_grad=True)
-            workspace = torch.zeros(q.size(), device='xpu', dtype=dtype, requires_grad=True)
-            grad_q_tmp = torch.zeros(q.size(), device='xpu', dtype=dtype, requires_grad=True)
-            alpha = sm_scale
-            dropout_prob = 0
-            grad_query = torch.empty_like(q, device='xpu', dtype=dtype, requires_grad=True)
-            grad_key = torch.empty_like(k, device='xpu', dtype=dtype, requires_grad=True)
-            grad_value = torch.empty_like(v, device='xpu', dtype=dtype, requires_grad=True)
-            grad_bias = torch.empty_like(bias, device='xpu', dtype=dtype, requires_grad=True)
-            bias_strideB = -1
-            bias_strideN = -1
-            bias_strideF = -1
-            attn_mask_padding = 0
-
-            xetla_fn = lambda: func(grad_out, q, k, v, bias, dropout, out, log_sumexp, workspace, grad_q_tmp, alpha,
-                                    dropout_prob, grad_query, grad_key, grad_value, grad_bias, Z, H, D_HEAD, N_CTX,
-                                    N_CTX, bias_strideB, bias_strideN, bias_strideF, attn_mask_padding)
-        _, min_ms, max_ms, mean, cv = benchmark_suit.do_bench(xetla_fn, n_warmup=10, n_repeat=10, quantiles=quantiles)
+    elif provider == 'onednn':
+        # OneDNN only supports MHA.
+        if H_q == H_kv:
+            mask = create_mask(causal_mask, 1, 1, N_CTX_q, N_CTX_kv, device=q.device)
+            xformers_fn = lambda: F.scaled_dot_product_attention(q, k, v, attn_mask=mask)
+            if MODE == 'bwd':
+                xformers_o = xformers_fn()
+                xformers_do = torch.randn_like(xformers_o)
+                xformers_fn = lambda: xformers_o.backward(xformers_do, retain_graph=True)
+            _, min_ms, max_ms, mean, cv = benchmark_suit.do_bench(xformers_fn, n_warmup=10, n_repeat=10,
+                                                                  quantiles=quantiles)
+        else:
+            _, min_ms, max_ms, mean, cv = float('nan'), float('nan'), float('nan'), float('nan'), float('nan')
 
     else:
         raise NotImplementedError(f'Unsupported provider {provider}')
 
-    tflops = lambda mean: 2 * 2 * Z * H * N_CTX * N_CTX * D_HEAD * (1e-12) / (mean * 1e-3)
-    gbps = lambda mean: Z * H * (N_CTX * D_HEAD + N_CTX * D_HEAD) * 2 * 2 * (1e-9) / (mean * 1e-3)
+    qk_flops = H_q * N_CTX_q * N_CTX_kv * D_HEAD_qk * 2  # mul + add
+    pv_flops = H_q * N_CTX_q * D_HEAD_v * N_CTX_kv * 2  # mul + add
+    tflops = lambda mean: Z * (qk_flops + pv_flops) * (1e-12) / (mean * 1e-3)
+
+    q_elems = H_q * N_CTX_q * D_HEAD_qk
+    k_elems = H_kv * N_CTX_kv * D_HEAD_qk
+    v_elems = H_kv * N_CTX_kv * D_HEAD_v
+    gbps = lambda mean: Z * (q_elems + k_elems + v_elems) * 2 * (1e-9) / (mean * 1e-3)  # float16 2 bytes
 
     if MODE == 'bwd':
-        tflops = lambda mean: 2.5 * 2 * 2 * Z * H * N_CTX * N_CTX * D_HEAD * (1e-12) / (mean * 1e-3)
-        gbps = lambda mean: 2.5 * Z * H * (N_CTX * D_HEAD + N_CTX * D_HEAD) * 2 * 2 * (1e-9) / (mean * 1e-3)
+        tflops = lambda mean: 2.5 * 2 * 2 * Z * H_q * N_CTX_q * N_CTX_kv * D_HEAD_qk * (1e-12) / (mean * 1e-3)
+        gbps = lambda mean: 2.5 * Z * H_q * (N_CTX_q * D_HEAD_qk + N_CTX_kv * D_HEAD_qk) * 2 * 2 * (1e-9) / (mean * 1e-3
+                                                                                                             )
 
     return (gbps(mean), gbps(max_ms), gbps(min_ms)), (tflops(mean), tflops(max_ms), tflops(min_ms)), cv
 
