[FlexFlash] Wire up mask_mod + blockmask to flash impl (#166359)

I have some local changes that I need to push to flash first
https://github.com/Dao-AILab/flash-attention/pull/1970

Pull Request resolved: https://github.com/pytorch/pytorch/pull/166359
Approved by: https://github.com/v0i0

test/inductor/test_flex_flash.py

@@ -122,16 +122,52 @@ def cuda_kernel_profiler(kernel_pattern="flash_attncute"):
     result["found"] = any(kernel_pattern in name for name in kernel_names)
 
 
-def flash_vs_triton(q, k, v, score_mod=None, rtol=5e-3, atol=5e-3):
+def flash_vs_triton(q, k, v, score_mod=None, block_mask=None, rtol=2):
     compiled_fn = torch.compile(flex_attention)
+
+    out_ref_fp32 = flex_attention(
+        q.to(torch.float32),
+        k.to(torch.float32),
+        v.to(torch.float32),
+        score_mod=score_mod,
+        block_mask=block_mask,
+    ).to(q.dtype)
+
     out_flash = compiled_fn(
-        q, k, v, score_mod=score_mod, kernel_options={"force_flash": True}
+        q,
+        k,
+        v,
+        score_mod=score_mod,
+        block_mask=block_mask,
+        kernel_options={"force_flash": True},
     )
-    out_no_flash = compiled_fn(
-        q, k, v, score_mod=score_mod, kernel_options={"force_flash": False}
+    out_triton = compiled_fn(
+        q,
+        k,
+        v,
+        score_mod=score_mod,
+        block_mask=block_mask,
+        kernel_options={"force_flash": False},
     )
-    torch.testing.assert_close(out_flash, out_no_flash, rtol=rtol, atol=atol)
-    return out_flash, out_no_flash
+
+    assert out_flash.shape == out_ref_fp32.shape == out_triton.shape
+    assert not torch.isnan(out_flash).any()
+    assert not torch.isnan(out_triton).any()
+    assert not torch.isnan(out_ref_fp32).any()
+    assert torch.isfinite(out_flash).all()
+    assert torch.isfinite(out_triton).all()
+    assert torch.isfinite(out_ref_fp32).all()
+
+    fwd_atol = 2 * (out_ref_fp32 + 0.3 - 0.3 - out_ref_fp32).abs().max().item()
+
+    triton_error = (out_triton - out_ref_fp32).abs().max().item()
+    flash_error = (out_flash - out_ref_fp32).abs().max().item()
+
+    assert flash_error <= rtol * triton_error + fwd_atol, (
+        f"Flash error {flash_error:.2e} exceeds {rtol}x Triton error {triton_error:.2e} + {fwd_atol:.2e}"
+    )
+
+    return out_flash, out_triton, out_ref_fp32
 
 
 def name_fn(score_mod):
@@ -162,26 +198,6 @@ class TestFlexFlash(InductorTestCase):
         q, k, v = create_test_tensors(seq_len=seq_len, dtype=dtype, device=device)
         flash_vs_triton(q, k, v, score_mod=_causal)
 
-    @dtypes(torch.float16, torch.bfloat16)
-    def test_force_flash_error_with_block_mask(self, device, dtype):
-        """Test that force_flash=True raises error when BlockMask is provided."""
-        q, k, v = create_test_tensors(dtype=dtype, device=device)
-
-        # Create a causal block mask
-        def causal_mask(b, h, q_idx, kv_idx):
-            return q_idx >= kv_idx
-
-        block_mask = create_block_mask(causal_mask, 2, 4, 512, 512, device=device)
-
-        compiled_fn = torch.compile(flex_attention)
-        with self.assertRaisesRegex(
-            RuntimeError,
-            r"force_flash=True but flash attention cannot be used.*BlockMask.*not supported",
-        ):
-            compiled_fn(
-                q, k, v, block_mask=block_mask, kernel_options={"force_flash": True}
-            )
-
     @dtypes(torch.float16, torch.bfloat16)
     def test_flash_attention_kernel_called(self, device, dtype):
         """Test that flash attention kernel is actually called when force_flash=True."""
@@ -257,7 +273,6 @@ class TestFlexFlash(InductorTestCase):
         """Test that force_flash=True raises error when tensor requires gradients."""
         q, k, v = create_test_tensors(dtype=dtype, device=device)
 
-        # Create a score mod with requires_grad tensor
         bias = torch.randn(4, device=device, dtype=dtype, requires_grad=True)
 
         def score_mod_with_grad(score, b, h, q_idx, kv_idx):
@@ -276,6 +291,108 @@ class TestFlexFlash(InductorTestCase):
                 kernel_options={"force_flash": True},
             )
 
+    @dtypes(torch.float16, torch.bfloat16)
+    def test_flash_attention_with_block_mask(self, device, dtype):
+        """Test flash attention with block mask and mask_mod."""
+        q, k, v = create_test_tensors(dtype=dtype, device=device)
+
+        def causal_mask(b, h, q_idx, kv_idx):
+            return q_idx >= kv_idx
+
+        block_mask = create_block_mask(causal_mask, 2, 4, 512, 512, device=device)
+        flash_vs_triton(q, k, v, block_mask=block_mask)
+
+    @dtypes(torch.float16, torch.bfloat16)
+    def test_flash_attention_block_mask_with_score_mod(self, device, dtype):
+        """Test flash attention with both block mask and score_mod."""
+        q, k, v = create_test_tensors(dtype=dtype, device=device)
+
+        def causal_mask(b, h, q_idx, kv_idx):
+            return q_idx >= kv_idx
+
+        block_mask = create_block_mask(causal_mask, 2, 4, 512, 512, device=device)
+        flash_vs_triton(q, k, v, score_mod=_times_two, block_mask=block_mask)
+
+    @dtypes(torch.float16, torch.bfloat16)
+    def test_flash_attention_with_mask_mod_buffer(self, device, dtype):
+        """Test flash attention with mask_mod that loads from buffer."""
+        q, k, v = create_test_tensors(
+            batch_size=2, num_heads=4, dtype=dtype, device=device
+        )
+
+        mask_bias = torch.randn(4, device=device, dtype=dtype) * 0.1
+
+        def custom_mask(b, h, q_idx, kv_idx):
+            bias_value = mask_bias[h]
+            return (q_idx >= kv_idx) | (bias_value > 0)
+
+        block_mask = create_block_mask(custom_mask, 2, 4, 512, 512, device=device)
+        flash_vs_triton(q, k, v, block_mask=block_mask)
+
+    @dtypes(torch.float16, torch.bfloat16)
+    def test_flash_attention_mask_mod_with_dual_buffers(self, device, dtype):
+        """Mask modifier should support multiple captured buffers."""
+        batch_size, num_heads, seq_len = 2, 4, 512
+        q, k, v = create_test_tensors(
+            batch_size=batch_size, num_heads=num_heads, dtype=dtype, device=device
+        )
+
+        head_bias = torch.randn(num_heads, device=device, dtype=dtype) * 0.2
+        batch_bias = torch.randn(batch_size, device=device, dtype=dtype) * 0.2
+
+        def dual_buffer_mask(b, h, q_idx, kv_idx):
+            head_term = head_bias[h]
+            batch_term = batch_bias[b]
+            causal = q_idx >= kv_idx
+            bias_cond = (head_term + batch_term).to(torch.float32) > 0
+            return causal | bias_cond
+
+        block_mask = create_block_mask(
+            dual_buffer_mask, batch_size, num_heads, seq_len, seq_len, device=device
+        )
+        flash_vs_triton(q, k, v, block_mask=block_mask)
+
+    @dtypes(torch.float16, torch.bfloat16)
+    def test_flash_attention_score_mod_with_many_buffer_indexing(self, device, dtype):
+        batch_size, num_heads, seq_len = 2, 4, 512
+        q, k, v = create_test_tensors(
+            batch_size=batch_size, num_heads=num_heads, dtype=dtype, device=device
+        )
+
+        head_bias = torch.randn(num_heads, device=device, dtype=dtype) * 0.15
+        query_scale = torch.randn(seq_len, device=device, dtype=dtype) * 0.05
+        kv_scale = torch.randn(seq_len, device=device, dtype=dtype) * 0.05
+        batch_bias = torch.randn(batch_size, device=device, dtype=dtype) * 0.1
+
+        def complex_score(score, b, h, q_idx, kv_idx):
+            head_term = head_bias[h]
+            query_term = query_scale[q_idx]
+            kv_term = kv_scale[kv_idx]
+            batch_term = batch_bias[b]
+            return score + head_term + query_term - kv_term + batch_term
+
+        flash_vs_triton(q, k, v, score_mod=complex_score)
+
+    @dtypes(torch.float16, torch.bfloat16)
+    def test_flash_attention_with_score_and_mask_buffers(self, device, dtype):
+        """Test flash attention with both score_mod and mask_mod using buffers."""
+        q, k, v = create_test_tensors(
+            batch_size=2, num_heads=4, dtype=dtype, device=device
+        )
+
+        score_bias = torch.randn(4, device=device, dtype=dtype) * 0.2
+        mask_bias = torch.randn(4, device=device, dtype=dtype) * 0.1
+
+        def score_with_buffer(score, b, h, q_idx, kv_idx):
+            return score + score_bias[h]
+
+        def mask_with_buffer(b, h, q_idx, kv_idx):
+            bias_value = mask_bias[h]
+            return (q_idx >= kv_idx) | (bias_value > 0)
+
+        block_mask = create_block_mask(mask_with_buffer, 2, 4, 512, 512, device=device)
+        flash_vs_triton(q, k, v, score_mod=score_with_buffer, block_mask=block_mask)
+
 
 instantiate_device_type_tests(TestFlexFlash, globals(), only_for="cuda")
 

torch/_inductor/codegen/cutedsl/cutedsl_kernel.py

@@ -65,6 +65,10 @@ class CuteDSLSubgraphInfo:
     body: IndentedBuffer
     template_mask: Optional[str] = None
     template_out: Optional[str] = None
+    cse: Optional[CSE[Any]] = None
+
+    def __post_init__(self):
+        self.only_copy_if_non_none_fields = ("cse",)
 
     def to_dict(self):
         return {
@@ -191,10 +195,15 @@ class CuteDSLTemplateKernel(Kernel):
                 body=IndentedBuffer(),
                 template_mask=None,
                 template_out=None,
+                cse=None,
             )
 
         subgraph = self.subgraph_bodies[body_name]
         for key, value in subgraph.to_dict().items():
+            if value is None and key in getattr(
+                subgraph, "only_copy_if_non_none_fields", ()
+            ):
+                continue
             setattr(self, key, value)
 
         try:
@@ -212,15 +221,17 @@ class CuteDSLTemplateKernel(Kernel):
                 setattr(self, key, value)
 
     @contextlib.contextmanager
-    def create_subgraph_body(self, body_name: str):
+    def create_subgraph_body(self, body_name: str, *, clear_cse: bool = False):
         """Create a new subgraph body for template processing."""
         assert body_name not in self.subgraph_bodies, (
             f"Subgraph body '{body_name}' already exists"
         )
+        new_cse = self.cse.clone() if clear_cse else None
         self.subgraph_bodies[body_name] = CuteDSLSubgraphInfo(
             body=IndentedBuffer(),
             template_mask=None,
             template_out=None,
+            cse=new_cse,
         )
         with self.set_subgraph_body(body_name):
             yield
@@ -294,7 +305,8 @@ class CuteDSLTemplateKernel(Kernel):
 
         # Register the hook and return placeholder
         placeholder = "<UNPACK_BUFFERS>"
-        assert placeholder not in self.render_hooks
+        # TODO: I think double invoking is fine for this specific hook
+        # assert placeholder not in self.render_hooks
         self.render_hooks[placeholder] = hook
         return placeholder
 
@@ -330,7 +342,7 @@ class CuteDSLTemplateKernel(Kernel):
         while f"mod_{subgraph_number}_{num}" in self.subgraph_bodies:
             num += 1
 
-        with self.create_subgraph_body(f"mod_{subgraph_number}_{num}"):
+        with self.create_subgraph_body(f"mod_{subgraph_number}_{num}", clear_cse=True):
             subgraph = self._get_subgraph(subgraph_number)
             modification_handler = ModificationWrapperCuteDSL(
                 self, subgraph_number, fixed_inputs, mask
@@ -429,40 +441,20 @@ class ModificationWrapperCuteDSL(V.WrapperHandler):  # type: ignore[name-defined
             #   val_frag[0] = tensor[index]
             #   result = val_frag.load()
 
-            index_frag = self.kernel.cse.generate(
-                self.kernel.body,
-                "cute.make_fragment(1, cutlass.Int32)",
-                dtype=torch.int32,
-                bounds=ValueRanges.unknown(),
+            index_frag = self.kernel.cse.newvar(dtype=torch.int32)
+            self.kernel.body.writeline(
+                f"{index_frag} = cute.make_fragment(1, cutlass.Int32)"
+            )
+            self.kernel.body.writeline(f"{index_frag}.store({index_str})")
+
+            val_frag = self.kernel.cse.newvar(dtype=var_dtype)
+            self.kernel.body.writeline(
+                f"{val_frag} = cute.make_fragment(1, {cute_dtype})"
             )
 
-            self.kernel.cse.generate(
-                self.kernel.body,
-                f"{index_frag}.store({index_str})",
-                dtype=torch.int32,
-                bounds=ValueRanges.unknown(),
-            )
-
-            val_frag = self.kernel.cse.generate(
-                self.kernel.body,
-                f"cute.make_fragment(1, {cute_dtype})",
-                dtype=var_dtype,
-                bounds=ValueRanges.unknown(),
-            )
-
-            index_var = self.kernel.cse.generate(
-                self.kernel.body,
-                f"{index_frag}[0]",
-                dtype=torch.int32,
-                bounds=ValueRanges.unknown(),
-            )
-
-            self.kernel.cse.generate(
-                self.kernel.body,
-                f"{val_frag}[0] = ({var}[{index_var}])",
-                dtype=var_dtype,
-                bounds=ValueRanges.unknown(),
-            )
+            index_var = self.kernel.cse.newvar(dtype=torch.int32)
+            self.kernel.body.writeline(f"{index_var} = {index_frag}[0]")
+            self.kernel.body.writeline(f"{val_frag}[0] = ({var}[{index_var}])")
 
             final_expr = f"{val_frag}.load()"
 

torch/_inductor/kernel/flex/flex_attention.py

@@ -193,24 +193,6 @@ def flex_attention(
             score_mod_other_buffers,
             mask_mod_other_buffers,
         )
-    if _use_flex_flash_attention(
-        subgraph,
-        mask_graph,
-        kernel_options,
-        num_score_mod_placeholders=len(placeholder_inps),
-    ):
-        return create_flex_flash_attention_kernel(
-            query,
-            key,
-            value,
-            block_mask,
-            scale,
-            kernel_options,
-            subgraph_buffer,
-            mask_graph_buffer,
-            score_mod_other_buffers,
-            mask_mod_other_buffers,
-        )
 
     (
         query,
@@ -240,6 +222,30 @@ def flex_attention(
         ]
     )
 
+    if _use_flex_flash_attention(
+        subgraph,
+        mask_graph,
+        kernel_options,
+        num_score_mod_placeholders=len(placeholder_inps),
+    ):
+        return create_flex_flash_attention_kernel(
+            query,
+            key,
+            value,
+            block_mask,
+            scale,
+            kernel_options,
+            subgraph_buffer,
+            mask_graph_buffer,
+            score_mod_other_buffers,
+            mask_mod_other_buffers,
+            kv_num_blocks,
+            kv_indices,
+            full_kv_num_blocks,
+            full_kv_indices,
+            mask_graph=mask_graph,
+        )
+
     score_mod_other_buffers = maybe_realize(score_mod_other_buffers)
     mask_mod_other_buffers = maybe_realize(mask_mod_other_buffers)
 

torch/_inductor/kernel/flex/flex_flash_attention.py

@@ -56,10 +56,8 @@ def input_buffers_require_grads(graph_module, num_score_mod_placeholders: int):
     return any(requires_grad(n) for n in inputs[num_score_mod_placeholders:])
 
 
-def is_trivial_graph(
-    graph_module: GraphModule, is_score_graph: bool, num_score_mod_placeholders: int
-):
-    """Check if the flex graphs are compatible with Flash Attention."""
+def is_trivial_mask_graph(graph_module: GraphModule) -> bool:
+    """Mask graph is trivial when it only gates via the default full op."""
     graph = graph_module.graph
     nodes = list(graph.nodes)
     placeholders = [n for n in nodes if n.op == "placeholder"]
@@ -67,14 +65,16 @@ def is_trivial_graph(
     assert len(output) == 1, "Got graph w/ multiple outputs"
     output_val = output[0].args[0]
 
-    if is_score_graph:
-        if input_buffers_require_grads(graph_module, num_score_mod_placeholders):
-            return False
-        return True  # party on garth
     # mask mod graph is empty if we have 4 inputs and full_default output
     return len(placeholders) == 4 and output_val.target is torch.ops.aten.full.default
 
 
+@functools.lru_cache(maxsize=1)
+def _supports_nontrivial_mask_graphs() -> bool:
+    """Currently only supported on Hopper (SM90) GPUs."""
+    return torch.cuda.get_device_capability()[0] == 9
+
+
 def _can_use_flex_flash_attention(
     subgraph: Subgraph, mask_graph: Subgraph, num_score_mod_placeholders: int
 ) -> tuple[bool, str]:
@@ -91,32 +91,15 @@ def _can_use_flex_flash_attention(
             False,
             "Input buffers require gradients (not supported by flash attention)",
         )
+    mask_trivial = is_trivial_mask_graph(mask_graph.graph_module)
 
-    score_trivial = is_trivial_graph(
-        subgraph.graph_module,
-        is_score_graph=True,
-        num_score_mod_placeholders=num_score_mod_placeholders,
-    )
-    mask_trivial = is_trivial_graph(
-        mask_graph.graph_module,
-        is_score_graph=False,
-        num_score_mod_placeholders=num_score_mod_placeholders,
-    )
+    if mask_trivial:
+        return True, ""
 
-    if not score_trivial and not mask_trivial:
+    if not _supports_nontrivial_mask_graphs():
         return (
             False,
-            "Both score and mask graphs are too complex for flash attention (require simple operations only)",
-        )
-    elif not score_trivial:
-        return (
-            False,
-            "Score modification captured tensors that require gradients (not supported by flash attention)",
-        )
-    elif not mask_trivial:
-        return (
-            False,
-            "A non None BlockMask was passed to flex attention (not supported by flash attention yet)",
+            "NYI: Non-trivial mask graphs only supported on Hopper (SM90) for flash attention",
         )
 
     return True, ""
@@ -154,6 +137,11 @@ def create_flex_flash_attention_kernel(
     mask_graph_buffer: SubgraphResults,
     score_mod_other_buffers: list[TensorBox],
     mask_mod_other_buffers: list[TensorBox],
+    kv_num_blocks: TensorBox | None,
+    kv_indices: TensorBox | None,
+    full_kv_num_blocks: TensorBox | None,
+    full_kv_indices: TensorBox | None,
+    mask_graph: Subgraph,
 ) -> tuple[TensorBox | ShapeAsConstantBuffer, TensorBox | ShapeAsConstantBuffer]:
     """Create a flex flash attention kernel using CuteDSL template."""
     if not ensure_flash_available():
@@ -193,17 +181,34 @@ def create_flex_flash_attention_kernel(
         stride=[sympy.sympify(s) for s in output.get_stride()],
     )
 
+    # Used to check if we can skip block sparse impl
+    mask_graph_is_trivial = is_trivial_mask_graph(mask_graph.graph_module)
+
+    needs_block_mask = not mask_graph_is_trivial
+    has_full_blocks = full_kv_num_blocks is not None
+
     choices: list[Any] = []
-    causal = kernel_options.get("causal", False)
     assert flash_attention_cutedsl_template is not None
+
+    input_nodes = [query, key, value, lse]
+    if has_full_blocks:
+        input_nodes.extend(
+            [kv_num_blocks, kv_indices, full_kv_num_blocks, full_kv_indices]
+        )
+
+    if needs_block_mask and not has_full_blocks:
+        raise NotImplementedError(
+            "Flash attention with block mask but without full blocks is not supported yet"
+        )
+
     error = flash_attention_cutedsl_template.maybe_append_choice(
         choices,
-        input_nodes=[query, key, value, lse],
+        input_nodes=input_nodes,
         layout=output_layout,
         mutated_inputs=[lse],
         subgraphs=[subgraph_buffer, mask_graph_buffer],
         SM_SCALE=scale,
-        CAUSAL=causal,
+        NEEDS_BLOCK_MASK=needs_block_mask,
     )
 
     if error or not choices:

torch/_inductor/kernel/flex/templates/flash_attention.py.jinja

@@ -1,6 +1,10 @@
-
+{% if NEEDS_BLOCK_MASK %}
+{{def_kernel("Q", "K", "V", "LOGSUMEXP", "KV_NUM_BLKS", "KV_IDX", "FULL_KV_NUM_BLKS", "FULL_KV_IDX")}}
+{% else %}
 {{def_kernel("Q", "K", "V", "LOGSUMEXP")}}
+{% endif %}
     from flash_attn.cute.interface import _flash_attn_fwd
+    from flash_attn.cute.block_sparsity import BlockSparseTensorsTorch
 
     # Transpose tensors for _flash_attn_fwd compatibility (B,H,M,D) -> (B,M,H,D)
     q_transposed = Q.transpose(1, 2)
@@ -26,6 +30,25 @@
     output = {{get_output()}}
     output_transposed = output.transpose(1, 2)
 
+    {% if NEEDS_BLOCK_MASK %}
+    @cute.jit
+    def mask_mod(b_idx, h_idx, q_idx, kv_idx, aux_tensors):
+        {{unpack_buffers("aux_tensors", indent_width=8)}}
+        {{ modification(
+        subgraph_number=1,
+        output_name="mask_mod_output",
+        b="b_idx",
+        h="h_idx",
+        m="q_idx",
+        n="kv_idx",
+        ) | indent_except_first(2) }}
+        return mask_mod_output
+    block_sparse_tensors = BlockSparseTensorsTorch(KV_NUM_BLKS, KV_IDX, FULL_KV_NUM_BLKS, FULL_KV_IDX)
+    {% else %}
+    block_sparse_tensors = None
+    mask_mod = None
+    {% endif %}
+
     # Collect any additional tensor buffers that were added during modifications
     {% set tensor_buffers = get_tensor_buffers() -%}
     {% if tensor_buffers -%}
@@ -41,10 +64,11 @@
         k_transposed,
         v_transposed,
         softmax_scale={{SM_SCALE}},
-        causal={{CAUSAL}},
         return_lse=True,
         score_mod=score_mod,
+        mask_mod=mask_mod,
         out=output_transposed,
         lse=LOGSUMEXP,
+        block_sparse_tensors=block_sparse_tensors,
         aux_tensors=buffers
-    )
+    )