fix lint

Author: YunchaoYang

src/fairseq2/models/qwen/__init__.py

@@ -12,7 +12,9 @@
 from fairseq2.models.qwen.config import Qwen35Config as Qwen35Config
 from fairseq2.models.qwen.config import Qwen35MoeConfig as Qwen35MoeConfig
 from fairseq2.models.qwen.config import QwenConfig as QwenConfig
-from fairseq2.models.qwen.config import register_qwen35_configs as register_qwen35_configs
+from fairseq2.models.qwen.config import (
+    register_qwen35_configs as register_qwen35_configs,
+)
 from fairseq2.models.qwen.config import (
     register_qwen35_moe_configs as register_qwen35_moe_configs,
 )
@@ -32,7 +34,9 @@
 from fairseq2.models.qwen.hub import (
     get_qwen35_moe_tokenizer_hub as get_qwen35_moe_tokenizer_hub,
 )
-from fairseq2.models.qwen.hub import get_qwen35_tokenizer_hub as get_qwen35_tokenizer_hub
+from fairseq2.models.qwen.hub import (
+    get_qwen35_tokenizer_hub as get_qwen35_tokenizer_hub,
+)
 from fairseq2.models.qwen.hub import get_qwen_model_hub as get_qwen_model_hub
 from fairseq2.models.qwen.hub import get_qwen_tokenizer_hub as get_qwen_tokenizer_hub
 from fairseq2.models.qwen.interop import (

src/fairseq2/models/qwen/attention.py

@@ -89,22 +89,30 @@ def __init__(
         # Q projection is DOUBLED — half query, half gate.
         # HF: nn.Linear(hidden, num_heads * head_dim * 2, bias=False)
         self.q_proj = Linear(
-            model_dim, num_heads * head_dim * 2, bias=False,
+            model_dim,
+            num_heads * head_dim * 2,
+            bias=False,
             init_fn=qkv_proj_init_fn,
         )
 
         self.k_proj = Linear(
-            model_dim, num_key_value_heads * head_dim, bias=False,
+            model_dim,
+            num_key_value_heads * head_dim,
+            bias=False,
             init_fn=qkv_proj_init_fn,
         )
 
         self.v_proj = Linear(
-            model_dim, num_key_value_heads * head_dim, bias=False,
+            model_dim,
+            num_key_value_heads * head_dim,
+            bias=False,
             init_fn=qkv_proj_init_fn,
         )
 
         self.output_proj = Linear(
-            num_heads * head_dim, model_dim, bias=False,
+            num_heads * head_dim,
+            model_dim,
+            bias=False,
             init_fn=output_proj_init_fn,
         )
 
@@ -163,12 +171,8 @@ def forward(
                 k_rot = k[..., :encoding_dim]
                 k_pass = k[..., encoding_dim:]
 
-                q_rot = self.pos_encoder(
-                    q_rot, seqs_layout, state_bag=state_bag
-                )
-                k_rot = self.pos_encoder(
-                    k_rot, keys_layout, state_bag=state_bag
-                )
+                q_rot = self.pos_encoder(q_rot, seqs_layout, state_bag=state_bag)
+                k_rot = self.pos_encoder(k_rot, keys_layout, state_bag=state_bag)
 
                 q = torch.cat([q_rot, q_pass], dim=-1)
                 k = torch.cat([k_rot, k_pass], dim=-1)
@@ -198,9 +202,7 @@ def forward(
 
         # -- Scaled dot-product attention --
         # q, k, v: (B, S, H, D)
-        attn_output, _ = self.sdpa(
-            q, seqs_layout, k, keys_layout, v, bias_cache
-        )
+        attn_output, _ = self.sdpa(q, seqs_layout, k, keys_layout, v, bias_cache)
 
         # -- Output gating --
         # attn_output: (B, S, H, D) -> (B, S, H * D)

src/fairseq2/models/qwen/config.py

@@ -65,6 +65,8 @@ class QwenConfig:
 
     pad_idx: int | None = None
     """The index of the pad symbol in the vocabulary."""
+
+
 # ---------------------------------------------------------------------------
 
 

src/fairseq2/models/qwen/decoder_layer.py

@@ -114,9 +114,7 @@ def forward(
         *,
         state_bag: IncrementalStateBag | None = None,
     ) -> Tensor:
-        seqs = self._forward_token_mixer(
-            seqs, seqs_layout, attn_bias_cache, state_bag
-        )
+        seqs = self._forward_token_mixer(seqs, seqs_layout, attn_bias_cache, state_bag)
         seqs = self._forward_ffn(seqs)
         return seqs
 

src/fairseq2/models/qwen/interop.py

@@ -47,6 +47,7 @@ def convert_qwen_state_dict(
 
     return state_dict
 
+
 # HG-side RMSNorm key suffixes for reverse conversion (weight -= 1.0).
 _QWEN35_HG_RMSNORM_SUFFIXES = (
     "input_layernorm.weight",
@@ -287,7 +288,7 @@ def _expand_with_language_model_prefix(
 # Components not yet integrated in the text-only CausalLM model.
 _QWEN35_VL_SKIP_PREFIXES: Final = (
     "model.visual.",  # vision encoder
-    "mtp.",           # multi-token prediction head
+    "mtp.",  # multi-token prediction head
 )