Add fp8_attention example and unit test

yf225 · yf225 · commit ea04d03e4b51 · 2025-07-14T12:55:04.000-07:00
diff --git a/examples/fp8_attention.py b/examples/fp8_attention.py
@@ -0,0 +1,233 @@
+from __future__ import annotations
+
+import math
+
+import torch
+
+import helion
+import helion.language as hl
+
+
+@helion.kernel(static_shapes=True)
+def fp8_attention_kernel(
+    q: torch.Tensor,  # [batch*heads, seq, dim]
+    k: torch.Tensor,  # [batch*heads, seq, dim]
+    v: torch.Tensor,  # [batch*heads, dim, seq] - pre-transposed
+) -> torch.Tensor:
+    batch_heads = q.size(0)
+    seq_len = q.size(1)
+    head_dim = q.size(2)
+
+    # Output tensor
+    out = torch.empty(
+        [batch_heads, seq_len, head_dim], dtype=torch.float32, device=q.device
+    )
+
+    # Scale factor for attention
+    sm_scale = 1.0 / math.sqrt(float(head_dim))
+    # Triton kernel multiplies sm_scale by 1.44269504 (1/log(2)) for exp2
+    sm_scale = sm_scale * 1.44269504
+
+    # Process each batch*head in parallel
+    for bh in hl.grid(batch_heads):
+        # Process each query position
+        for tile_m in hl.tile(seq_len):
+            # Initialize for online softmax
+            m_i = hl.full([tile_m], float("-inf"), dtype=torch.float32)
+            l_i = hl.full([tile_m], 0.0, dtype=torch.float32)
+            acc = hl.zeros([tile_m, head_dim], dtype=torch.float32)
+
+            # Load query tile - keep in FP8
+            q_tile = q[bh, tile_m, :]  # [tile_m, dim]
+
+            # Compute attention scores for all keys
+            for tile_n in hl.tile(seq_len):
+                # Load key tile and transpose for Q @ K^T
+                k_tile = k[bh, tile_n, :]  # [tile_n, dim] - keep in FP8
+                k_tile_t = k_tile.transpose(0, 1)  # [dim, tile_n]
+
+                # Compute Q @ K^T with FP8 inputs, result in FP32
+                qk = torch.matmul(q_tile, k_tile_t).to(
+                    torch.float32
+                )  # [tile_m, tile_n]
+
+                # Scale QK scores first
+                qk_scaled = qk * sm_scale  # [tile_m, tile_n]
+
+                # Compute max of scaled scores
+                qk_max = torch.amax(qk_scaled, dim=-1)  # [tile_m]
+
+                # Update global max
+                m_new = torch.maximum(m_i, qk_max)
+
+                # Shift by max for numerical stability
+                qk_shifted = qk_scaled - m_new[:, None]
+
+                # Use exp2 to match Triton kernel's implementation
+                # Note: Triton kernel already multiplies sm_scale by 1.44269504
+                p = torch.exp2(qk_shifted)  # [tile_m, tile_n]
+
+                # Sum of exponentials for this block
+                l_ij = torch.sum(p, dim=-1)  # [tile_m]
+
+                # Update accumulators with correction factor
+                # Correction factor for previous blocks
+                alpha = torch.exp2(m_i - m_new)
+                l_i = l_i * alpha + l_ij
+                acc = acc * alpha[:, None]
+
+                # Load values - V is [dim, seq]
+                v_tile = v[bh, :, tile_n]  # [dim, tile_n] - keep in FP8
+
+                # Convert p to FP8 for FP8 GEMM
+                p_fp8 = p.to(v.dtype)  # Convert to same FP8 type as V
+
+                # Accumulate attention @ V with FP8 GEMM
+                v_t = v_tile.transpose(0, 1)  # [tile_n, dim]
+                pv = torch.matmul(p_fp8, v_t).to(torch.float32)  # [tile_m, dim]
+                acc = acc + pv
+
+                # Update max tracker
+                m_i = m_new
+
+            # Final normalization
+            acc = acc / l_i[:, None]
+            out[bh, tile_m, :] = acc
+
+    return out
+
+
+def prepare_fp8_attention_inputs(
+    q: torch.Tensor,  # [batch, heads, seq, dim]
+    k: torch.Tensor,  # [batch, heads, seq, dim]
+    v: torch.Tensor,  # [batch, heads, seq, dim]
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, tuple[int, int, int, int]]:
+    """
+    Common preprocessing for FP8 attention implementations.
+
+    Returns:
+        q_reshaped_fp8: [batch*heads, seq, dim] - in FP8 e5m2
+        k_reshaped_fp8: [batch*heads, seq, dim] - in FP8 e5m2
+        v_transposed_fp8: [batch*heads, dim, seq] - in FP8 e5m2
+        shape: (batch, heads, seq_len, head_dim)
+    """
+    batch, heads, seq_len, head_dim = q.shape
+
+    # Reshape to [batch*heads, seq, dim]
+    q_reshaped = q.reshape(batch * heads, seq_len, head_dim)
+    k_reshaped = k.reshape(batch * heads, seq_len, head_dim)
+
+    # Transpose V to [batch, heads, dim, seq] then reshape
+    v_transposed = v.permute(0, 1, 3, 2).reshape(batch * heads, head_dim, seq_len)
+
+    # Convert to FP8 e5m2
+    q_reshaped_fp8 = q_reshaped.to(torch.float8_e5m2)
+    k_reshaped_fp8 = k_reshaped.to(torch.float8_e5m2)
+    v_transposed_fp8 = v_transposed.to(torch.float8_e5m2)
+
+    return (
+        q_reshaped_fp8,
+        k_reshaped_fp8,
+        v_transposed_fp8,
+        (batch, heads, seq_len, head_dim),
+    )
+
+
+def fp8_attention_tritonbench(
+    q: torch.Tensor, k: torch.Tensor, v: torch.Tensor
+) -> torch.Tensor:
+    """Wrapper for TritonBench compatibility."""
+    # Common preprocessing with FP8 conversion
+    q_fp8, k_fp8, v_fp8, shape = prepare_fp8_attention_inputs(q, k, v)
+    batch, heads, seq_len, head_dim = shape
+
+    # Call the fused kernel
+    out_fused = fp8_attention_kernel(q_fp8, k_fp8, v_fp8)
+
+    # Reshape back and convert to FP16
+    out = out_fused.reshape(batch, heads, seq_len, head_dim)
+    return out.to(torch.float16)
+
+
+def fp8_attention_pytorch(
+    q: torch.Tensor,  # [batch, heads, seq, dim]
+    k: torch.Tensor,  # [batch, heads, seq, dim]
+    v: torch.Tensor,  # [batch, heads, seq, dim]
+) -> torch.Tensor:
+    """
+    Baseline PyTorch implementation of FP8 attention using FP8 e5m2.
+    """
+    # Get preprocessed inputs with FP8 conversion
+    q_fp8, k_fp8, v_fp8, shape = prepare_fp8_attention_inputs(q, k, v)
+    batch, heads, seq_len, head_dim = shape
+
+    sm_scale = 1.0 / math.sqrt(float(head_dim))
+
+    outputs = []
+
+    for i in range(batch * heads):
+        q_i = q_fp8[i]  # [seq, dim] - already FP8
+        k_i = k_fp8[i]  # [seq, dim] - already FP8
+        v_i = v_fp8[i]  # [dim, seq] - pre-transposed, already FP8
+
+        # For Q @ K^T, we need K^T to be column-major
+        kt_fp8 = k_i.t()  # column-major [dim, seq]
+
+        # Q @ K^T - dequantize and use regular matmul since e5m2 not supported by _scaled_mm
+        q_deq = q_i.to(torch.float32)
+        kt_deq = kt_fp8.to(torch.float32)
+        qk = torch.matmul(q_deq, kt_deq)
+
+        # Compute max before scaling
+        qk_max = torch.amax(qk, dim=-1, keepdim=True)
+
+        # Scale and shift in one operation, then use exp2
+        qk_scaled_shifted = qk * sm_scale - qk_max * sm_scale
+        p = torch.exp2(qk_scaled_shifted * 1.44269504)
+
+        # Normalize
+        p_norm = p / p.sum(dim=-1, keepdim=True)
+
+        # Step 2: Attention @ V using FP8
+        # P is [seq, seq], V is [dim, seq]
+        # We want P @ V^T = [seq, seq] @ [seq, dim] = [seq, dim]
+        p_fp8 = p_norm.to(torch.float8_e5m2)  # row-major [seq, seq]
+
+        # v_i is [dim, seq], already FP8
+        vt_fp8 = v_i.t()  # column-major [seq, dim]
+
+        # P @ V^T - dequantize and use regular matmul since e5m2 not supported by torch._scaled_mm
+        p_deq = p_fp8.to(torch.float32)
+        vt_deq = vt_fp8.to(torch.float32)
+        out_i = torch.matmul(p_deq, vt_deq)
+
+        outputs.append(out_i)
+
+    # Stack and reshape back
+    out_stacked = torch.stack(outputs, dim=0)  # [batch*heads, seq, dim]
+    out = out_stacked.reshape(batch, heads, seq_len, head_dim)
+
+    return out.to(torch.float16)
+
+
+def check(batch: int, heads: int, seq_len: int, head_dim: int) -> None:
+    torch.manual_seed(42)
+    q = torch.randn(batch, heads, seq_len, head_dim, dtype=torch.float16, device="cuda")
+    k = torch.randn(batch, heads, seq_len, head_dim, dtype=torch.float16, device="cuda")
+    v = torch.randn(batch, heads, seq_len, head_dim, dtype=torch.float16, device="cuda")
+
+    from helion._testing import run_example
+
+    run_example(
+        fp8_attention_tritonbench, fp8_attention_pytorch, (q, k, v), atol=0.1, rtol=0.1
+    )
+
+
+def main() -> None:
+    check(1, 2, 128, 64)
+    check(2, 4, 256, 64)
+    check(4, 8, 512, 128)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/test/test_examples.expected b/test/test_examples.expected
@@ -535,6 +535,94 @@ def _embedding_make_precompiler(x: torch.Tensor, weight: torch.Tensor):
     from helion.runtime.precompile_shim import make_precompiler
     return make_precompiler(_embedding_kernel)(x_flat, weight, out, x_flat.size(0), out.stride(0), out.stride(1), weight.stride(0), weight.stride(1), x_flat.stride(0), embedding_dim, _BLOCK_SIZE_1, num_warps=4, num_stages=3)
 
+--- assertExpectedJournal(TestExamples.test_fp8_attention)
+from __future__ import annotations
+
+import math
+import torch
+import triton
+import triton.language as tl
+from torch._inductor.runtime import triton_helpers
+from torch._inductor.runtime.triton_compat import libdevice
+
+@triton.jit
+def _fp8_attention_kernel_kernel(q, k, v, out, _RDIM_SIZE_2: tl.constexpr, _BLOCK_SIZE_1: tl.constexpr, _BLOCK_SIZE_3: tl.constexpr):
+    pid_0 = tl.program_id(0)
+    offset_0 = pid_0
+    indices_5 = tl.arange(0, _RDIM_SIZE_2).to(tl.int32)
+    for offset_4 in tl.range(0, 256, _BLOCK_SIZE_1):
+        indices_4 = offset_4 + tl.arange(0, _BLOCK_SIZE_1).to(tl.int32)
+        m_i = tl.full([_BLOCK_SIZE_1], float('-inf'), tl.float32)
+        l_i = tl.full([_BLOCK_SIZE_1], 0.0, tl.float32)
+        acc = tl.full([_BLOCK_SIZE_1, 64], 0.0, tl.float32)
+        q_tile = tl.load(q + (offset_0 * 16384 + indices_4[:, None] * 64 + indices_5[None, :] * 1), None)
+        for offset_2 in tl.range(0, 256, _BLOCK_SIZE_3):
+            indices_2 = offset_2 + tl.arange(0, _BLOCK_SIZE_3).to(tl.int32)
+            q_tile_copy = q_tile
+            m_i_copy = m_i
+            l_i_copy = l_i
+            acc_copy = acc
+            q_tile_copy_0 = q_tile_copy
+            m_i_copy_0 = m_i_copy
+            l_i_copy_0 = l_i_copy
+            acc_copy_0 = acc_copy
+            k_tile = tl.load(k + (offset_0 * 16384 + indices_2[:, None] * 64 + indices_5[None, :] * 1), None)
+            k_tile_t = tl.permute(k_tile, [1, 0])
+            mm = tl.dot(q_tile_copy_0, k_tile_t)
+            v_0 = mm.to(tl.float32)
+            v_1 = 0.18033688
+            v_2 = v_0 * v_1
+            qk_max = tl.max(v_2, 1)
+            v_3 = triton_helpers.maximum(m_i_copy_0, qk_max)
+            subscript = v_3[:, None]
+            v_4 = v_2 - subscript
+            v_5 = libdevice.exp2(v_4)
+            l_ij = tl.sum(v_5, 1)
+            v_6 = m_i_copy_0 - v_3
+            v_7 = libdevice.exp2(v_6)
+            v_8 = l_i_copy_0 * v_7
+            l_i = v_8 + l_ij
+            subscript_1 = v_7[:, None]
+            v_10 = acc_copy_0 * subscript_1
+            v_tile = tl.load(v + (offset_0 * 16384 + indices_5[:, None] * 1 + indices_2[None, :] * 64), None)
+            v_11 = v_5.to(tl.float8e5)
+            v_t = tl.permute(v_tile, [1, 0])
+            mm_1 = tl.dot(v_11, v_t)
+            v_12 = mm_1.to(tl.float32)
+            acc = v_10 + v_12
+            m_i = v_3
+        subscript_2 = l_i[:, None]
+        v_14 = acc / subscript_2
+        tl.store(out + (offset_0 * 16384 + indices_4[:, None] * 64 + indices_5[None, :] * 1), v_14, None)
+
+def fp8_attention_kernel(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
+    """FP8 attention kernel processing batch*heads in parallel."""
+    batch_heads = q.size(0)
+    seq_len = q.size(1)
+    head_dim = q.size(2)
+    out = torch.empty([batch_heads, seq_len, head_dim], dtype=torch.float32, device=q.device)
+    sm_scale = 1.0 / math.sqrt(float(head_dim))
+    sm_scale = sm_scale * 1.44269504
+    _RDIM_SIZE_2 = 64
+    _BLOCK_SIZE_1 = 64
+    _BLOCK_SIZE_3 = 64
+    _fp8_attention_kernel_kernel[8,](q, k, v, out, _RDIM_SIZE_2, _BLOCK_SIZE_1, _BLOCK_SIZE_3, num_warps=4, num_stages=3)
+    return out
+
+def _fp8_attention_kernel_make_precompiler(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):
+    """FP8 attention kernel processing batch*heads in parallel."""
+    batch_heads = q.size(0)
+    seq_len = q.size(1)
+    head_dim = q.size(2)
+    out = torch.empty([batch_heads, seq_len, head_dim], dtype=torch.float32, device=q.device)
+    sm_scale = 1.0 / math.sqrt(float(head_dim))
+    sm_scale = sm_scale * 1.44269504
+    _RDIM_SIZE_2 = 64
+    _BLOCK_SIZE_1 = 64
+    _BLOCK_SIZE_3 = 64
+    from helion.runtime.precompile_shim import make_precompiler
+    return make_precompiler(_fp8_attention_kernel_kernel)(q, k, v, out, _RDIM_SIZE_2, _BLOCK_SIZE_1, _BLOCK_SIZE_3, num_warps=4, num_stages=3)
+
 --- assertExpectedJournal(TestExamples.test_fp8_gemm)
 from __future__ import annotations
 
diff --git a/test/test_examples.py b/test/test_examples.py
@@ -523,6 +523,47 @@ def test_segment_reduction(self):
             )
         )
 
+    @unittest.skipIf(
+        not torch.cuda.is_available() or torch.cuda.get_device_capability()[0] < 9,
+        "FP8 requires GPU with compute capability >= 9.0 (e.g., H100)",
+    )
+    def test_fp8_attention(self):
+        batch = 2
+        heads = 4
+        seq_len = 256
+        head_dim = 64
+
+        # Create FP16 tensors
+        q = torch.randn(
+            batch, heads, seq_len, head_dim, dtype=torch.float16, device=DEVICE
+        )
+        k = torch.randn(
+            batch, heads, seq_len, head_dim, dtype=torch.float16, device=DEVICE
+        )
+        v = torch.randn(
+            batch, heads, seq_len, head_dim, dtype=torch.float16, device=DEVICE
+        )
+
+        # Import the module
+        mod = import_path(EXAMPLES_DIR / "fp8_attention.py")
+
+        # Prepare FP8 inputs using the module's preprocessing function
+        q_fp8, k_fp8, v_fp8, shape = mod.prepare_fp8_attention_inputs(q, k, v)
+        args = (q_fp8, k_fp8, v_fp8)
+
+        # Get expected output from kernel
+        expected = mod.fp8_attention_kernel(*args)
+
+        self.assertExpectedJournal(
+            check_example(
+                "fp8_attention",
+                args,
+                expected,
+                fn_name="fp8_attention_kernel",
+                block_sizes=[64, 64],
+            )
+        )
+
 
 if __name__ == "__main__":
     unittest.main()
