Bump to v0.2.6

Dao-AILab · Dec 28, 2022 · a6ec178 · a6ec178
1 parent 63670fd
commit a6ec178
Show file tree

Hide file tree

Showing 3 changed files with 3 additions and 3 deletions.
diff --git a/flash_attn/modules/mha.py b/flash_attn/modules/mha.py
@@ -436,7 +436,7 @@ def forward(self, x, x_kv=None, key_padding_mask=None, cu_seqlens=None, max_seql
                 kv = self._update_kv_cache(qkv[:, :, 1:], inference_params)
                 # If we're processing the prompt, causal=None (use self.causal).
                 # If we're decoding, then causal=False.
-                causal = False if inference_params.sequence_len_offset == 0 else None
+                causal = None if inference_params.sequence_len_offset == 0 else False
                 context = self.inner_cross_attn(q, kv, causal=causal)
         else:
             if not self.return_residual:

diff --git a/flash_attn/utils/generation.py b/flash_attn/utils/generation.py
@@ -40,7 +40,7 @@ def greedy_decode(input_ids, model, max_length):
         inference_params.sequence_len_offset = seqlen_og
         while True:
             position_ids = torch.full((batch_size, 1), inference_params.sequence_len_offset,
-                                    dtype=torch.long, device=input_ids.device)
+                                      dtype=torch.long, device=input_ids.device)
             logits = model(rearrange(next_token, 'b -> b 1'), position_ids=position_ids,
                            inference_params=inference_params).logits[:, -1]
             scores.append(logits)

diff --git a/setup.py b/setup.py
@@ -156,7 +156,7 @@ def append_nvcc_threads(nvcc_extra_args):
 
 setup(
     name="flash_attn",
-    version="0.2.5",
+    version="0.2.6-1",
     packages=find_packages(
         exclude=("build", "csrc", "include", "tests", "dist", "docs", "benchmarks", "flash_attn.egg-info",)
     ),