[Lint] Phaseout Yapf format and embrace ruff format (#1417)

examples/bitnet-1.58b/vllm_workspace/utils.py

@@ -3,8 +3,7 @@ from typing import Dict, List, Tuple
 TokensText = Tuple[List[int], str]
 
 
-def check_outputs_equal(outputs_0_lst: List[TokensText], outputs_1_lst: List[TokensText],
-                        name_0: str, name_1: str):
+def check_outputs_equal(outputs_0_lst: List[TokensText], outputs_1_lst: List[TokensText], name_0: str, name_1: str):
     """
     Compare the two sequences generated by different models,
     which should be equal.
@@ -15,19 +14,14 @@ def check_outputs_equal(outputs_0_lst: List[TokensText], outputs_1_lst: List[Tok
         output_ids_0, output_str_0 = outputs_0
         output_ids_1, output_str_1 = outputs_1
 
-        assert output_str_0 == output_str_1, (f"Test{prompt_idx}:"
-                                              f"\n{name_0}:\t{output_str_0!r}"
-                                              f"\n{name_1}:\t{output_str_1!r}")
-        assert output_ids_0 == output_ids_1, (f"Test{prompt_idx}:"
-                                              f"\n{name_0}:\t{output_str_0!r}"
-                                              f"\n{name_1}:\t{output_str_1!r}")
+        assert output_str_0 == output_str_1, f"Test{prompt_idx}:\n{name_0}:\t{output_str_0!r}\n{name_1}:\t{output_str_1!r}"
+        assert output_ids_0 == output_ids_1, f"Test{prompt_idx}:\n{name_0}:\t{output_str_0!r}\n{name_1}:\t{output_str_1!r}"
 
 
 TokensTextLogprobs = Tuple[List[int], str, List[Dict[int, float]]]
 
 
-def check_logprobs_close(outputs_0_lst: List[TokensTextLogprobs],
-                         outputs_1_lst: List[TokensTextLogprobs], name_0: str, name_1: str):
+def check_logprobs_close(outputs_0_lst: List[TokensTextLogprobs], outputs_1_lst: List[TokensTextLogprobs], name_0: str, name_1: str):
     """
     Compare the logprobs of two sequences generated by different models,
     which should be similar but not necessarily equal.
@@ -41,16 +35,11 @@ def check_logprobs_close(outputs_0_lst: List[TokensTextLogprobs],
 
         # Loop through generated tokens.
         for idx, (output_id_0, output_id_1) in enumerate(zip(output_ids_0, output_ids_1)):
-
             # If generated tokens don't match, then
             if output_id_0 != output_id_1:
                 # Each predicted token must be in top N logprobs of the other
-                assert output_id_0 in logprobs_1[idx], (f"Test{prompt_idx}:"
-                                                        f"\n{name_0}:\t{output_str_0!r}"
-                                                        f"\n{name_1}:\t{output_str_1!r}")
-                assert output_id_1 in logprobs_0[idx], (f"Test{prompt_idx}:"
-                                                        f"\n{name_0}:\t{output_str_0!r}"
-                                                        f"\n{name_1}:\t{output_str_1!r}")
+                assert output_id_0 in logprobs_1[idx], f"Test{prompt_idx}:\n{name_0}:\t{output_str_0!r}\n{name_1}:\t{output_str_1!r}"
+                assert output_id_1 in logprobs_0[idx], f"Test{prompt_idx}:\n{name_0}:\t{output_str_0!r}\n{name_1}:\t{output_str_1!r}"
 
                 # Break out since sequences will now diverge.
                 break

examples/blocksparse_attention/block_sparse_attn_triton.py

@@ -15,10 +15,7 @@ def get_sparse_attn_mask_from_topk(x, topk, use_dense_for_last_block=False):
     bsz, num_head, downsample_len, _ = x.shape
     # N_CTX = downsample_len * BLOCK
     sparse_index = torch.topk(x, topk, dim=-1).indices
-    dense_mask = torch.full([bsz, num_head, downsample_len, downsample_len],
-                            False,
-                            dtype=torch.bool,
-                            device=x.device)
+    dense_mask = torch.full([bsz, num_head, downsample_len, downsample_len], False, dtype=torch.bool, device=x.device)
     dense_mask.scatter_(-1, sparse_index, True)
     if use_dense_for_last_block:
         dense_mask[:, :, -2:, :] = True
@@ -56,7 +53,6 @@ def _fwd_kernel_inner(
     BLOCK_M: tl.constexpr,
     BLOCK_N: tl.constexpr,
 ):
-
     mask_val = tl.load(block_mask_ptr + k_block_col_idx * stride_bmask_n)
     # print
 
@@ -73,8 +69,7 @@ def _fwd_kernel_inner(
 
         # the following is needed only when LAST_K_BLOCK or BLOCK_M < BLOCK_N
         if LAST_K_BLOCK:
-            qk += tl.where(offs_m[:, None] + past_len >= (start_n + offs_n[None, :]), 0,
-                           float('-inf'))
+            qk += tl.where(offs_m[:, None] + past_len >= (start_n + offs_n[None, :]), 0, float("-inf"))
 
         m_ij = tl.maximum(m_i, tl.max(qk, 1))
         qk -= m_ij[:, None]
@@ -154,7 +149,7 @@ def _fwd_kernel(
     v_ptrs = V + off_v
     mask_ptrs = block_mask_ptr + start_m * stride_bmm
 
-    m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float('inf')
+    m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
     l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
     acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
 
@@ -192,24 +187,12 @@ def _fwd_kernel(
     acc = acc * l_recip
     acc = acc.to(Out.dtype.element_ty)
 
-    off_o = off_z * stride_oz + off_h * stride_oh + offs_m[:, None] * stride_om + offs_d[
-        None, :] * stride_od
+    off_o = off_z * stride_oz + off_h * stride_oh + offs_m[:, None] * stride_om + offs_d[None, :] * stride_od
     out_ptrs = Out + off_o
     tl.store(out_ptrs, acc, mask=offs_m[:, None] < N_CTX)
 
 
-def _forward(ctx,
-             q,
-             k,
-             v,
-             block_sparse_mask,
-             sm_scale,
-             BLOCK_M=64,
-             BLOCK_N=64,
-             num_warps=None,
-             num_stages=1,
-             out=None):
-
+def _forward(ctx, q, k, v, block_sparse_mask, sm_scale, BLOCK_M=64, BLOCK_N=64, num_warps=None, num_stages=1, out=None):
     assert q.shape[-1] == k.shape[-1] == v.shape[-1]
     assert k.shape[2] == v.shape[2]
     o = out if out is not None else torch.empty_like(q).contiguous()
@@ -254,7 +237,6 @@ def _forward(ctx,
 
 
 class _sparse_attention(torch.autograd.Function):
-
     @staticmethod
     def forward(ctx, q, k, v, block_sparse_dense, sm_scale):
         # shape constraints
@@ -278,9 +260,9 @@ def test_topk_sparse_attention():
     torch.manual_seed(0)
 
     # Create inputs
-    q = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device='cuda', dtype=torch.bfloat16)
-    k = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device='cuda', dtype=torch.bfloat16)
-    v = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device='cuda', dtype=torch.bfloat16)
+    q = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device="cuda", dtype=torch.bfloat16)
+    k = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device="cuda", dtype=torch.bfloat16)
+    v = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device="cuda", dtype=torch.bfloat16)
     sm_scale = 1.0 / (D_HEAD**0.5)
 
     # Create sparse mask (downsampled to block level)
@@ -288,9 +270,7 @@ def test_topk_sparse_attention():
     downsample_len = math.ceil(SEQ_LEN / downsample_factor)
     print("downsample_len", downsample_len)
 
-    x_ds = torch.randn([BATCH, N_HEADS, downsample_len, downsample_len],
-                       device='cuda',
-                       dtype=torch.bfloat16)
+    x_ds = torch.randn([BATCH, N_HEADS, downsample_len, downsample_len], device="cuda", dtype=torch.bfloat16)
     x_ds[:, :, :, 0] = 100
     print("x_ds.shape", x_ds.shape)
     block_mask = get_sparse_attn_mask_from_topk(x_ds, topk=TOPK)
@@ -302,22 +282,21 @@ def test_topk_sparse_attention():
 
     # Compute reference
     # Expand block mask to full attention matrix
-    full_mask = torch.kron(block_mask.float(), torch.ones(BLOCK, BLOCK, device='cuda'))
+    full_mask = torch.kron(block_mask.float(), torch.ones(BLOCK, BLOCK, device="cuda"))
     full_mask = full_mask[..., :SEQ_LEN, :SEQ_LEN].bool()
     full_mask = full_mask & torch.tril(torch.ones_like(full_mask))  # Apply causal
 
     # PyTorch reference implementation
-    attn = torch.einsum('bhsd,bhtd->bhst', q, k) * sm_scale
-    attn = attn.masked_fill(~full_mask, float('-inf'))
+    attn = torch.einsum("bhsd,bhtd->bhst", q, k) * sm_scale
+    attn = attn.masked_fill(~full_mask, float("-inf"))
     attn = F.softmax(attn, dim=-1)
-    ref_output = torch.einsum('bhst,bhtd->bhsd', attn, v)
+    ref_output = torch.einsum("bhst,bhtd->bhsd", attn, v)
 
     # print("ref_output", ref_output)
     # print("triton_output", triton_output)
 
     # Verify accuracy
-    assert torch.allclose(triton_output, ref_output, atol=1e-2, rtol=1e-2), \
-        "Triton output doesn't match reference"
+    assert torch.allclose(triton_output, ref_output, atol=1e-2, rtol=1e-2), "Triton output doesn't match reference"
     print("Pass topk sparse attention test with qlen == klen")
 
 
@@ -329,9 +308,9 @@ def test_topk_sparse_attention_qlt_kl():
     torch.manual_seed(0)
 
     # Create inputs.
-    q = torch.randn(BATCH, N_HEADS, Q_LEN, D_HEAD, device='cuda', dtype=torch.bfloat16)
-    k = torch.randn(BATCH, N_HEADS, K_LEN, D_HEAD, device='cuda', dtype=torch.bfloat16)
-    v = torch.randn(BATCH, N_HEADS, K_LEN, D_HEAD, device='cuda', dtype=torch.bfloat16)
+    q = torch.randn(BATCH, N_HEADS, Q_LEN, D_HEAD, device="cuda", dtype=torch.bfloat16)
+    k = torch.randn(BATCH, N_HEADS, K_LEN, D_HEAD, device="cuda", dtype=torch.bfloat16)
+    v = torch.randn(BATCH, N_HEADS, K_LEN, D_HEAD, device="cuda", dtype=torch.bfloat16)
     # softmax scale
     sm_scale = 1.0 / (D_HEAD**0.5)
 
@@ -339,8 +318,7 @@ def test_topk_sparse_attention_qlt_kl():
     print("downsample_factor", downsample_factor)
     downsample_len = math.ceil(K_LEN / downsample_factor)  # number of blocks along one dimension
     print("downsample_len", downsample_len)
-    x_ds = torch.randn(
-        BATCH, N_HEADS, downsample_len, downsample_len, device='cuda', dtype=torch.bfloat16)
+    x_ds = torch.randn(BATCH, N_HEADS, downsample_len, downsample_len, device="cuda", dtype=torch.bfloat16)
     # Force the first column to be high so that the first block is always selected.
     x_ds[:, :, :, 0] = 100
     block_mask = get_sparse_attn_mask_from_topk(x_ds, topk=TOPK)
@@ -351,26 +329,25 @@ def test_topk_sparse_attention_qlt_kl():
 
     past_len = K_LEN - Q_LEN
 
-    attn = torch.einsum('bhsd,bhtd->bhst', q, k) * sm_scale
+    attn = torch.einsum("bhsd,bhtd->bhst", q, k) * sm_scale
 
-    full_mask_full = torch.kron(block_mask.float(), torch.ones(BLOCK, BLOCK, device='cuda')).bool()
+    full_mask_full = torch.kron(block_mask.float(), torch.ones(BLOCK, BLOCK, device="cuda")).bool()
     full_mask_full = full_mask_full[..., :K_LEN, :K_LEN]
 
     effective_mask = full_mask_full[..., past_len:K_LEN, :]  # shape: (B, H, Q_LEN, K_LEN)
 
     i_global = torch.arange(past_len, K_LEN, device=k.device).unsqueeze(1)  # shape: (Q_LEN, 1)
     j_global = torch.arange(K_LEN, device=k.device).unsqueeze(0)  # shape: (1, K_LEN)
-    causal_mask = (j_global <= i_global)  # shape: (Q_LEN, K_LEN)
+    causal_mask = j_global <= i_global  # shape: (Q_LEN, K_LEN)
 
     final_mask = effective_mask & causal_mask  # shape: (B, H, Q_LEN, K_LEN)
 
-    attn = attn.masked_fill(~final_mask, float('-inf'))
+    attn = attn.masked_fill(~final_mask, float("-inf"))
     attn = F.softmax(attn, dim=-1)
-    ref_output = torch.einsum('bhst,bhtd->bhsd', attn, v)
+    ref_output = torch.einsum("bhst,bhtd->bhsd", attn, v)
 
     # Verify accuracy.
-    assert torch.allclose(triton_output, ref_output, atol=1e-2, rtol=1e-2), \
-        "Triton output doesn't match reference when qlen < klen"
+    assert torch.allclose(triton_output, ref_output, atol=1e-2, rtol=1e-2), "Triton output doesn't match reference when qlen < klen"
 
     print("Pass topk sparse attention test with qlen < klen")
 

examples/blocksparse_attention/example_tilelang_block_sparse_attn.py

@@ -10,10 +10,7 @@ def get_sparse_attn_mask_from_topk(x, topk, use_dense_for_last_block=False):
     bsz, num_head, downsample_len, _ = x.shape
     # N_CTX = downsample_len * BLOCK
     sparse_index = torch.topk(x, topk, dim=-1).indices
-    dense_mask = torch.full([bsz, num_head, downsample_len, downsample_len],
-                            False,
-                            dtype=torch.bool,
-                            device=x.device)
+    dense_mask = torch.full([bsz, num_head, downsample_len, downsample_len], False, dtype=torch.bool, device=x.device)
     dense_mask.scatter_(-1, sparse_index, True)
     if use_dense_for_last_block:
         dense_mask[:, :, -2:, :] = True
@@ -30,15 +27,17 @@ def get_sparse_attn_mask_from_threshold(x, threshold, use_dense_for_last_block=F
 
 
 @tilelang.jit(
-    out_idx=[4], pass_configs={
+    out_idx=[4],
+    pass_configs={
         tilelang.PassConfigKey.TL_ENABLE_FAST_MATH: True,
-    })
+    },
+)
 def blocksparse_flashattn(batch, heads, seq_len, dim, downsample_len, is_causal):
     block_M = 64
     block_N = 64
     num_stages = 1
     threads = 128
-    scale = (1.0 / dim)**0.5 * 1.44269504  # log2(e)
+    scale = (1.0 / dim) ** 0.5 * 1.44269504  # log2(e)
     shape = [batch, heads, seq_len, dim]
     block_mask_shape = [batch, heads, downsample_len, downsample_len]
 
@@ -47,7 +46,6 @@ def blocksparse_flashattn(batch, heads, seq_len, dim, downsample_len, is_causal)
     block_mask_dtype = "bool"
 
     def kernel_func(block_M, block_N, num_stages, threads):
-
         @T.macro
         def MMA0(
             K: T.Tensor(shape, dtype),
@@ -59,11 +57,10 @@ def blocksparse_flashattn(batch, heads, seq_len, dim, downsample_len, is_causal)
             by: T.int32,
             bz: T.int32,
         ):
-            T.copy(K[bz, by, k * block_N:(k + 1) * block_N, :], K_shared)
+            T.copy(K[bz, by, k * block_N : (k + 1) * block_N, :], K_shared)
             if is_causal:
                 for i, j in T.Parallel(block_M, block_N):
-                    acc_s[i, j] = T.if_then_else(bx * block_M + i >= k * block_N + j, 0,
-                                                 -T.infinity(acc_s.dtype))
+                    acc_s[i, j] = T.if_then_else(bx * block_M + i >= k * block_N + j, 0, -T.infinity(acc_s.dtype))
             else:
                 T.clear(acc_s)
             T.gemm(Q_shared, K_shared, acc_s, transpose_B=True, policy=T.GemmWarpPolicy.FullRow)
@@ -78,7 +75,7 @@ def blocksparse_flashattn(batch, heads, seq_len, dim, downsample_len, is_causal)
             by: T.int32,
             bz: T.int32,
         ):
-            T.copy(V[bz, by, k * block_N:(k + 1) * block_N, :], V_shared)
+            T.copy(V[bz, by, k * block_N : (k + 1) * block_N, :], V_shared)
             T.gemm(acc_s_cast, V_shared, acc_o, policy=T.GemmWarpPolicy.FullRow)
 
         @T.macro
@@ -127,8 +124,7 @@ def blocksparse_flashattn(batch, heads, seq_len, dim, downsample_len, is_causal)
             BlockSparseMask: T.Tensor(block_mask_shape, block_mask_dtype),
             Output: T.Tensor(shape, dtype),
         ):
-            with T.Kernel(
-                    T.ceildiv(seq_len, block_M), heads, batch, threads=threads) as (bx, by, bz):
+            with T.Kernel(T.ceildiv(seq_len, block_M), heads, batch, threads=threads) as (bx, by, bz):
                 Q_shared = T.alloc_shared([block_M, dim], dtype)
                 K_shared = T.alloc_shared([block_N, dim], dtype)
                 V_shared = T.alloc_shared([block_N, dim], dtype)
@@ -143,7 +139,7 @@ def blocksparse_flashattn(batch, heads, seq_len, dim, downsample_len, is_causal)
                 logsum = T.alloc_fragment([block_M], accum_dtype)
                 block_mask = T.alloc_local([downsample_len], block_mask_dtype)
 
-                T.copy(Q[bz, by, bx * block_M:(bx + 1) * block_M, :], Q_shared)
+                T.copy(Q[bz, by, bx * block_M : (bx + 1) * block_M, :], Q_shared)
                 T.fill(acc_o, 0)
                 T.fill(logsum, 0)
                 T.fill(scores_max, -T.infinity(accum_dtype))
@@ -152,20 +148,19 @@ def blocksparse_flashattn(batch, heads, seq_len, dim, downsample_len, is_causal)
                     block_mask[vj] = BlockSparseMask[bz, by, bx, vj]
 
                 loop_range = (
-                    T.min(T.ceildiv(seq_len, block_N), T.ceildiv(
-                        (bx + 1) * block_M, block_N)) if is_causal else T.ceildiv(seq_len, block_N))
+                    T.min(T.ceildiv(seq_len, block_N), T.ceildiv((bx + 1) * block_M, block_N)) if is_causal else T.ceildiv(seq_len, block_N)
+                )
 
                 for k in T.Pipelined(loop_range, num_stages=num_stages):
                     if block_mask[k] != 0:
                         MMA0(K, Q_shared, K_shared, acc_s, k, bx, by, bz)
-                        Softmax(acc_s, acc_s_cast, scores_max, scores_max_prev, scores_scale,
-                                scores_sum, logsum)
+                        Softmax(acc_s, acc_s_cast, scores_max, scores_max_prev, scores_scale, scores_sum, logsum)
                         Rescale(acc_o, scores_scale)
                         MMA1(V, V_shared, acc_s_cast, acc_o, k, by, bz)
                 for i, j in T.Parallel(block_M, dim):
                     acc_o[i, j] /= logsum[i]
                 T.copy(acc_o, O_shared)
-                T.copy(O_shared, Output[bz, by, bx * block_M:(bx + 1) * block_M, :])
+                T.copy(O_shared, Output[bz, by, bx * block_M : (bx + 1) * block_M, :])
 
         return blocksparse_flashattn
 
@@ -180,18 +175,16 @@ def test_topk_sparse_attention():
     torch.manual_seed(0)
 
     # Create inputs
-    q = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device='cuda', dtype=torch.float16)
-    k = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device='cuda', dtype=torch.float16)
-    v = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device='cuda', dtype=torch.float16)
+    q = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device="cuda", dtype=torch.float16)
+    k = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device="cuda", dtype=torch.float16)
+    v = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device="cuda", dtype=torch.float16)
 
     sm_scale = 1.0 / (D_HEAD**0.5)
 
     # Create sparse mask (downsampled to block level)
     downsample_factor = BLOCK
     downsample_len = math.ceil(SEQ_LEN / downsample_factor)
-    x_ds = torch.randn([BATCH, N_HEADS, downsample_len, downsample_len],
-                       device='cuda',
-                       dtype=torch.bfloat16)
+    x_ds = torch.randn([BATCH, N_HEADS, downsample_len, downsample_len], device="cuda", dtype=torch.bfloat16)
     x_ds[:, :, :, 0] = 100
     block_mask = get_sparse_attn_mask_from_topk(x_ds, topk=TOPK)
 
@@ -202,15 +195,15 @@ def test_topk_sparse_attention():
 
     # Compute reference
     # Expand block mask to full attention matrix
-    full_mask = torch.kron(block_mask.float(), torch.ones(BLOCK, BLOCK, device='cuda'))
+    full_mask = torch.kron(block_mask.float(), torch.ones(BLOCK, BLOCK, device="cuda"))
     full_mask = full_mask[..., :SEQ_LEN, :SEQ_LEN].bool()
     full_mask = full_mask & torch.tril(torch.ones_like(full_mask))  # Apply causal
 
     # PyTorch reference implementation
-    attn = torch.einsum('bhsd,bhtd->bhst', q, k) * sm_scale
-    attn = attn.masked_fill(~full_mask, float('-inf'))
+    attn = torch.einsum("bhsd,bhtd->bhst", q, k) * sm_scale
+    attn = attn.masked_fill(~full_mask, float("-inf"))
     attn = F.softmax(attn, dim=-1)
-    ref_output = torch.einsum('bhst,bhtd->bhsd', attn, v)
+    ref_output = torch.einsum("bhst,bhtd->bhsd", attn, v)
 
     print("ref_output", ref_output)
     print("tilelang_output", tilelang_output)

examples/blocksparse_attention/heuristic.py

 import math
 
 
-def num_splits_heuristic(total_mblocks, num_SMs, num_n_blocks, num_m_blocks, size_one_kv_head,
-                         is_causal_or_local, max_splits):
+def num_splits_heuristic(total_mblocks, num_SMs, num_n_blocks, num_m_blocks, size_one_kv_head, is_causal_or_local, max_splits):
     """
     Determines the optimal number of splits for maximizing GPU occupancy while balancing memory efficiency.
 

examples/blocksparse_attention/test_example_blocksparse_attention.py

@@ -25,26 +25,14 @@ def test_example_tilelang_sparse_gqa_decode_varlen_mask():
 
 def test_example_triton_sparse_gqa_decode_varlen_indice():
     example_triton_sparse_gqa_decode_varlen_indice.main(
-        batch=8,
-        heads=8,
-        heads_kv=4,
-        max_cache_seqlen=2048,
-        dim=128,
-        dim_v=128,
-        sparse_ratio=0.8,
-        block_size=32)
+        batch=8, heads=8, heads_kv=4, max_cache_seqlen=2048, dim=128, dim_v=128, sparse_ratio=0.8, block_size=32
+    )
 
 
 def test_example_triton_sparse_gqa_decode_varlen_mask():
     example_triton_sparse_gqa_decode_varlen_mask.main(
-        batch=16,
-        heads=16,
-        heads_kv=8,
-        max_cache_seqlen=1024,
-        dim=128,
-        dim_v=128,
-        sparse_ratio=0.8,
-        block_size=32)
+        batch=16, heads=16, heads_kv=8, max_cache_seqlen=1024, dim=128, dim_v=128, sparse_ratio=0.8, block_size=32
+    )
 
 
 if __name__ == "__main__":

examples/cast/example_triton_cast_to_fp8.py

@@ -128,9 +128,7 @@ def per_token_group_quant_fp8(
         Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
         scaling factor for quantization.
     """
-    assert (x.shape[-1] %
-            group_size == 0), (f"the last dimension of `x` {x.shape[-1]} must be divisible "
-                               f"by `group_size` {group_size}")
+    assert x.shape[-1] % group_size == 0, f"the last dimension of `x` {x.shape[-1]} must be divisible by `group_size` {group_size}"
     assert x.stride(-1) == 1, "`x` groups must be contiguous"
 
     finfo = torch.finfo(dtype)

examples/cast/test_example_cast.py

@@ -4,8 +4,7 @@ import example_per_token_cast_to_fp8
 
 
 def test_example_group_per_split_token_cast_to_fp8():
-    example_group_per_split_token_cast_to_fp8.main(
-        M=1024, N=1024, BG=2, blk_m=4, batch_sizes=[128, 896])
+    example_group_per_split_token_cast_to_fp8.main(M=1024, N=1024, BG=2, blk_m=4, batch_sizes=[128, 896])
 
 
 def test_example_per_token_cast_to_fp8():

examples/compile_flags/usecase.py

@@ -4,7 +4,6 @@ import tilelang.language as T
 
 # @tilelang.jit(compile_flags=["-O3", "--use_fast_math", "--expt-relaxed-constexpr"])
 def matmul(M, N, K, block_M, block_N, block_K, dtype="float16", accum_dtype="float"):
-
     @T.prim_func
     def main(
         A: T.Tensor((M, K), dtype),
@@ -36,8 +35,7 @@ block_K = 32
 
 func = matmul(M, N, K, block_M, block_N, block_K)
 
-jit_kernel = tilelang.compile(
-    func, out_idx=[2], target="cuda", compile_flags="-O3 --use_fast_math --expt-relaxed-constexpr")
+jit_kernel = tilelang.compile(func, out_idx=[2], target="cuda", compile_flags="-O3 --use_fast_math --expt-relaxed-constexpr")
 # or jit_kernel = tilelang.compile(func, out_idx=[2], target="cuda", compile_flags=["-O3", "--use_fast_math", "--expt-relaxed-constexpr"])
 # or jit_kernel = tilelang.compile(func, out_idx=[2], target="cuda", compile_flags=["-O3 --use_fast_math --expt-relaxed-constexpr"])