[Example] Specify a fixed commit for the flash-linear-attention repository and...

[Example] Specify a fixed commit for the flash-linear-attention repository and optimize nsa examples (#913)

- Updated the requirements.txt to specify a fixed commit for the flash-linear-attention repository.
- Refactored import paths in benchmark_nsa_fwd.py for better organization.
- Added a new function to generate configurations for autotuning.
- Modified the tilelang_sparse_attention function to accept parameters for block size, number of stages, and threads, enhancing flexibility.
- Changed allocation of shared memory for accumulators to optimize performance.

.gitignore

@@ -92,3 +92,6 @@ tilelang/jit/adapter/cython/.cycache
 
 # cache directory for clangd
 .cache/
+
+# claude
+**/.claude

examples/deepseek_nsa/benchmark/benchmark_nsa_fwd.py

@@ -10,7 +10,7 @@ from typing import Optional, Union
 from einops import rearrange, repeat
 import triton
 import triton.language as tl
-from fla.ops.common.utils import prepare_token_indices
+from fla.ops.utils import prepare_token_indices
 from fla.utils import autocast_custom_fwd, contiguous
 
 
@@ -439,6 +439,20 @@ def naive_nsa(q: torch.Tensor,
     return o_slc.to(dtype) + o_swa.to(dtype) if o_swa is not None else o_slc.to(dtype)
 
 
+def get_configs():
+    import itertools
+    iter_params = dict(
+        block_T=[128, 256, 512],
+        num_stages=[0, 1, 2, 4, 5],
+        threads=[32, 64, 128, 256, 512],
+    )
+    return [{
+        k: v for k, v in zip(iter_params, values)
+    } for values in itertools.product(*iter_params.values())]
+
+
+@tilelang.autotune(configs=get_configs(),)
+@tilelang.jit
 def tilelang_sparse_attention(batch,
                               heads,
                               seq_len,
@@ -447,7 +461,10 @@ def tilelang_sparse_attention(batch,
                               scale=None,
                               block_size=64,
                               groups=1,
-                              selected_blocks=16):
+                              selected_blocks=16,
+                              block_T=128,
+                              num_stages=2,
+                              threads=32):
     if scale is None:
         scale = (1.0 / dim)**0.5 * 1.44269504  # log2(e)
     else:
@@ -461,7 +478,7 @@ def tilelang_sparse_attention(batch,
     dtype = "float16"
     accum_dtype = "float"
     block_S = block_size
-    block_T = min(128, tilelang.math.next_power_of_2(dim))
+    block_T = min(block_T, tilelang.math.next_power_of_2(dim))
 
     NK = tilelang.cdiv(dim, block_T)
     NV = tilelang.cdiv(dim, block_T)
@@ -471,8 +488,6 @@ def tilelang_sparse_attention(batch,
     G = groups
     BS = block_S
     BK = BV = block_T
-    num_stages = 2
-    threads = 32
 
     @T.prim_func
     def tilelang_sparse_attention(
@@ -489,7 +504,7 @@ def tilelang_sparse_attention(batch,
             O_shared = T.alloc_shared([G, BV], dtype)
 
             acc_s = T.alloc_fragment([G, BS], accum_dtype)
-            acc_s_cast = T.alloc_fragment([G, BS], dtype)
+            acc_s_cast = T.alloc_shared([G, BS], dtype)
             acc_o = T.alloc_fragment([G, BV], accum_dtype)
             scores_max = T.alloc_fragment([G], accum_dtype)
             scores_max_prev = T.alloc_fragment([G], accum_dtype)
@@ -497,11 +512,7 @@ def tilelang_sparse_attention(batch,
             scores_sum = T.alloc_fragment([G], accum_dtype)
             logsum = T.alloc_fragment([G], accum_dtype)
 
-            # T.use_swizzle(10)
-
-            T.annotate_layout({Q_shared: tilelang.layout.make_swizzled_layout(Q_shared)})
-            T.annotate_layout({K_shared: tilelang.layout.make_swizzled_layout(K_shared)})
-            T.annotate_layout({V_shared: tilelang.layout.make_swizzled_layout(V_shared)})
+            T.annotate_layout({O_shared: tilelang.layout.make_swizzled_layout(O_shared)})
 
             i_t, i_v, i_bh = bx, by, bz
             i_b, i_h = i_bh // head_kv, i_bh % head_kv
@@ -597,7 +608,7 @@ def benchmark_nsa(batch_size,
     torch.random.manual_seed(0)
 
     # Compile the NSA kernel
-    program = tilelang_sparse_attention(
+    kernel = tilelang_sparse_attention(
         batch=batch_size,
         heads=head_query,
         seq_len=seq_len,
@@ -608,9 +619,6 @@ def benchmark_nsa(batch_size,
         selected_blocks=selected_blocks,
         scale=scale,
     )
-    print(program)
-    kernel = tilelang.compile(program, out_idx=None, execution_backend="cython")
-    print(kernel.get_kernel_source())
 
     profiler = kernel.get_profiler()
 

examples/deepseek_nsa/requirements.txt

-git+https://github.com/fla-org/flash-linear-attention
+git+https://github.com/fla-org/flash-linear-attention@c3bd56589033610264532b11f0972c69e4645f6e
\ No newline at end of file