Enable flake8-pie rules

pyproject.toml

@@ -32,7 +32,7 @@ line-length = 119
 ignore = ["C901", "E501", "E741", "F402", "F823", "SIM1", "SIM300", "SIM212", "SIM905", "UP009", "UP015", "UP031", "UP028", "UP004", "UP045", "UP007"]
 # RUF013: Checks for the use of implicit Optional
 #  in type annotations when the default parameter value is None.
-select = ["C", "E", "F", "I", "W", "RUF013", "PERF102", "PLC1802", "PLC0208", "SIM", "UP", "PIE794", "FURB"]
+select = ["C", "E", "F", "I", "W", "RUF013", "PERF102", "PLC1802", "PLC0208", "SIM", "UP", "PIE", "FURB"]
 extend-safe-fixes = ["UP006"]
 
 # Ignore import violations in all `__init__.py` files.

src/transformers/audio_utils.py

@@ -131,7 +131,7 @@ def load_audio_librosa(audio: Union[str, np.ndarray], sampling_rate=16000, timeo
     requires_backends(load_audio_librosa, ["librosa"])
 
     # Load audio from URL (e.g https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/translate_to_chinese.wav)
-    if audio.startswith("http://") or audio.startswith("https://"):
+    if audio.startswith(("http://", "https://")):
         audio = librosa.load(
             BytesIO(httpx.get(audio, follow_redirects=True, timeout=timeout).content), sr=sampling_rate
         )[0]

src/transformers/image_utils.py

@@ -456,7 +456,7 @@ def load_image(image: Union[str, "PIL.Image.Image"], timeout: Optional[float] =
     """
     requires_backends(load_image, ["vision"])
     if isinstance(image, str):
-        if image.startswith("http://") or image.startswith("https://"):
+        if image.startswith(("http://", "https://")):
             # We need to actually check for a real protocol, otherwise it's impossible to use a local file
             # like http_huggingface_co.png
             image = PIL.Image.open(BytesIO(httpx.get(image, timeout=timeout, follow_redirects=True).content))

src/transformers/integrations/finegrained_fp8.py

@@ -117,7 +117,7 @@ def _w8a8_block_fp8_matmul(
     Bs_ptrs = Bs + offs_bsn * stride_Bs_n
 
     accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
-    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
+    for k in range(tl.cdiv(K, BLOCK_SIZE_K)):
         a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0)
         b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)
 

src/transformers/modeling_gguf_pytorch_utils.py

@@ -120,7 +120,7 @@ def _split_moe_expert_tensor(
         # https://github.com/ggerganov/llama.cpp/blob/master/convert_hf_to_gguf.py#L1994-L2022
         name = tensor_key_mapping[name]
         w_counter = self.config.get("num_experts", 60)
-        for i in range(0, w_counter):
+        for i in range(w_counter):
             temp_name = name.replace("mlp.experts.", f"mlp.experts.{i}.")
             exp_weight = weights[i]
             parsed_parameters["tensors"][temp_name] = torch.from_numpy(np.copy(exp_weight))
@@ -338,7 +338,7 @@ def get_gguf_hf_weights_map(
             hf_name = re.sub(r"mlp.experts.\d+.", "mlp.experts.", hf_name)
 
         name, suffix = hf_name, ""
-        if hf_name.endswith(".weight") or hf_name.endswith(".bias"):
+        if hf_name.endswith((".weight", ".bias")):
             name, suffix = hf_name.rsplit(".", 1)
             suffix = "." + suffix
 

src/transformers/models/bert/convert_bert_token_dropping_original_tf2_checkpoint_to_pytorch.py

@@ -82,7 +82,7 @@ def get_encoder_attention_layer_array(layer_index: int, name: str, original_shap
     model = BertForMaskedLM(config)
 
     # Layers
-    for layer_index in range(0, config.num_hidden_layers):
+    for layer_index in range(config.num_hidden_layers):
         layer: BertLayer = model.bert.encoder.layer[layer_index]
 
         # Self-attention

src/transformers/models/bertweet/tokenization_bertweet.py

@@ -345,7 +345,7 @@ def normalizeToken(self, token):
         lowercased_token = token.lower()
         if token.startswith("@"):
             return "@USER"
-        elif lowercased_token.startswith("http") or lowercased_token.startswith("www"):
+        elif lowercased_token.startswith(("http", "www")):
             return "HTTPURL"
         elif len(token) == 1:
             if token in self.special_puncts:

src/transformers/models/big_bird/convert_bigbird_original_tf_checkpoint_to_pytorch.py

@@ -177,10 +177,8 @@ def load_tf_weights_trivia_qa(init_vars):
         try:
             if len(array.shape) > len(pointer.shape) and math.prod(array.shape) == math.prod(pointer.shape):
                 # print(txt_name, array.shape)
-                if (
-                    txt_name.endswith("attention/self/key/kernel")
-                    or txt_name.endswith("attention/self/query/kernel")
-                    or txt_name.endswith("attention/self/value/kernel")
+                if txt_name.endswith(
+                    ("attention/self/key/kernel", "attention/self/query/kernel", "attention/self/value/kernel")
                 ):
                     array = array.transpose(1, 0, 2).reshape(pointer.shape)
                 elif txt_name.endswith("attention/output/dense/kernel"):

src/transformers/models/clap/feature_extraction_clap.py

@@ -175,7 +175,7 @@ def _np_extract_fbank_features(self, waveform: np.ndarray, mel_filters: Optional
         return log_mel_spectrogram.T
 
     def _random_mel_fusion(self, mel, total_frames, chunk_frames):
-        ranges = np.array_split(list(range(0, total_frames - chunk_frames + 1)), 3)
+        ranges = np.array_split(list(range(total_frames - chunk_frames + 1)), 3)
         if len(ranges[1]) == 0:
             # if the audio is too short, we just use the first chunk
             ranges[1] = [0]

src/transformers/models/conditional_detr/convert_conditional_detr_original_pytorch_checkpoint_to_pytorch.py

@@ -279,7 +279,7 @@ def convert_conditional_detr_checkpoint(model_name, pytorch_dump_folder_path):
             elif "class_labels_classifier" in key or "bbox_predictor" in key:
                 val = state_dict.pop(key)
                 state_dict["conditional_detr." + key] = val
-            elif key.startswith("bbox_attention") or key.startswith("mask_head"):
+            elif key.startswith(("bbox_attention", "mask_head")):
                 continue
             else:
                 val = state_dict.pop(key)

src/transformers/models/deprecated/gptsan_japanese/convert_gptsan_tf_checkpoint_to_pytorch.py

@@ -40,7 +40,7 @@ def convert_tf_gptsan_to_pt(args):
         shapes = reader.get_variable_to_shape_map()
         for key_name in shapes:
             vnp = reader.get_tensor(key_name).astype(np.float16)
-            if key_name.endswith("/adam_m") or key_name.endswith("/adam_v"):
+            if key_name.endswith(("/adam_m", "/adam_v")):
                 continue
             if key_name.startswith("pasts/"):
                 if key_name.startswith("pasts/mlp"):
@@ -60,7 +60,7 @@ def convert_tf_gptsan_to_pt(args):
                     name = "model.blocks.%d.feed_forward.soft_bypass_mlp.weight" % player
                     state = vnp.transpose([1, 0]).copy()  # Mesh-Tensorflow is a diagonal matrix
                     new_state[name] = torch.tensor(state)
-                elif key_name.endswith("/wo/kernel") or key_name.endswith("/wi/kernel"):
+                elif key_name.endswith(("/wo/kernel", "/wi/kernel")):
                     nlayer = key_name[-9:-7]
                     for i in range(16):
                         name = "model.blocks.%d.feed_forward.mlp.experts.expert_%d.%s.weight" % (player, i, nlayer)
@@ -140,11 +140,7 @@ def convert_tf_gptsan_to_pt(args):
                     name = "model.blocks.%d.self_attn.norm.weight" % player
                     state = vnp.copy()  # same because it is one dimensional
                     new_state[name] = torch.tensor(state)
-            elif (
-                key_name.startswith("model/wte")
-                or key_name.startswith("model/wpe")
-                or key_name.startswith("model/ete")
-            ):
+            elif key_name.startswith(("model/wte", "model/wpe", "model/ete")):
                 nlayer = {"wte": "embed_tokens", "wpe": "position_embeddings", "ete": "extra_position_embeddings"}[
                     key_name[-3:]
                 ]

src/transformers/models/deprecated/jukebox/modeling_jukebox.py

@@ -96,7 +96,7 @@ def get_relevant_lyric_tokens(full_tokens, max_n_lyric_tokens, total_length, off
         tokens = torch.cat(
             [torch.zeros(max_n_lyric_tokens - len(full_tokens), dtype=torch.long).to(full_tokens.device), full_tokens]
         )
-        indices = [-1] * (max_n_lyric_tokens - len(full_tokens)) + list(range(0, len(full_tokens)))
+        indices = [-1] * (max_n_lyric_tokens - len(full_tokens)) + list(range(len(full_tokens)))
     else:
         midpoint = int(len(full_tokens) * (offset + duration / 2.0) / total_length)
         midpoint = min(max(midpoint, max_n_lyric_tokens // 2), len(full_tokens) - max_n_lyric_tokens // 2)
@@ -1464,7 +1464,7 @@ def sample(
             if get_preds:
                 preds = []
 
-            iter = tqdm(range(0, sample_tokens), leave=False)
+            iter = tqdm(range(sample_tokens), leave=False)
             for sample_t in iter:
                 iter.set_description(f"Ancestral sampling {sample_tokens} music tokens", refresh=True)
                 hidden_states, cond = self.get_emb(

src/transformers/models/detr/convert_detr_original_pytorch_checkpoint_to_pytorch.py

@@ -236,7 +236,7 @@ def convert_detr_checkpoint(model_name, pytorch_dump_folder_path):
             elif "class_labels_classifier" in key or "bbox_predictor" in key:
                 val = state_dict.pop(key)
                 state_dict["detr." + key] = val
-            elif key.startswith("bbox_attention") or key.startswith("mask_head"):
+            elif key.startswith(("bbox_attention", "mask_head")):
                 continue
             else:
                 val = state_dict.pop(key)

src/transformers/models/detr/convert_detr_to_pytorch.py

@@ -322,7 +322,7 @@ def convert_detr_checkpoint(model_name, pytorch_dump_folder_path=None, push_to_h
             elif "class_labels_classifier" in key or "bbox_predictor" in key:
                 val = state_dict.pop(key)
                 state_dict["detr." + key] = val
-            elif key.startswith("bbox_attention") or key.startswith("mask_head"):
+            elif key.startswith(("bbox_attention", "mask_head")):
                 continue
             else:
                 val = state_dict.pop(key)

src/transformers/models/funnel/modeling_funnel.py

@@ -152,7 +152,7 @@ def get_position_embeds(
             pos = torch.arange(0, seq_len, dtype=torch.int64, device=device).to(dtype)
             pooled_pos = pos
             position_embeds_list = []
-            for block_index in range(0, self.config.num_blocks):
+            for block_index in range(self.config.num_blocks):
                 # For each block with block_index > 0, we need two types position embeddings:
                 #   - Attention(pooled-q, unpooled-kv)
                 #   - Attention(pooled-q, pooled-kv)

src/transformers/models/glm4v/convert_glm4v_mgt_weights_to_hf.py

@@ -35,7 +35,7 @@ class DummyClass:
             def __init__(self, *args, **kwargs):
                 pass
 
-        if mod_name.startswith("megatron") or mod_name.startswith("glm") or mod_name.startswith("__main__"):
+        if mod_name.startswith(("megatron", "glm", "__main__")):
             return DummyClass
         return super().find_class(mod_name, name)
 

src/transformers/models/glm4v/modular_glm4v.py

@@ -1649,7 +1649,7 @@ def __call__(
                     timestamps = metadata.timestamps[::2]  # mrope
 
                     unique_timestamps = []
-                    for idx in range(0, len(timestamps)):
+                    for idx in range(len(timestamps)):
                         unique_timestamps.append(timestamps[idx])
 
                     selected_timestamps = unique_timestamps[:num_frames]

src/transformers/models/glm4v/processing_glm4v.py

@@ -177,7 +177,7 @@ def __call__(
                     timestamps = metadata.timestamps[::2]  # mrope
 
                     unique_timestamps = []
-                    for idx in range(0, len(timestamps)):
+                    for idx in range(len(timestamps)):
                         unique_timestamps.append(timestamps[idx])
 
                     selected_timestamps = unique_timestamps[:num_frames]

src/transformers/models/glm4v_moe/convert_glm4v_moe_mgt_weights_to_hf.py

@@ -30,7 +30,7 @@ class DummyClass:
             def __init__(self, *args, **kwargs):
                 pass
 
-        if mod_name.startswith("megatron") or mod_name.startswith("glm") or mod_name.startswith("__main__"):
+        if mod_name.startswith(("megatron", "glm", "__main__")):
             return DummyClass
         return super().find_class(mod_name, name)
 

src/transformers/models/gpt_neo/configuration_gpt_neo.py

@@ -178,7 +178,7 @@ def custom_unfold(input, dimension, size, step):
     s[dimension] = indices
     sliced = input[s]
 
-    perm = list(range(0, rank + 1))
+    perm = list(range(rank + 1))
     perm.append(perm.pop(dimension + 1))
 
     return sliced.permute(perm)

src/transformers/models/gpt_sw3/tokenization_gpt_sw3.py

@@ -135,7 +135,7 @@ def __init__(
 
         # Regular expression to remove non-printing characters (e.g. some unicode control chars) in preprocessing
         self.non_printing_characters_re = re.compile(
-            f"[{''.join(map(chr, list(range(0, 9)) + list(range(11, 32)) + list(range(127, 160)) + [160, 173, 8203]))}]"
+            f"[{''.join(map(chr, list(range(9)) + list(range(11, 32)) + list(range(127, 160)) + [160, 173, 8203]))}]"
         )
 
         super().__init__(

src/transformers/models/instructblipvideo/modeling_instructblipvideo.py

@@ -1230,7 +1230,6 @@ def get_image_features(
             pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
                 The tensors corresponding to the input images.
         """
-        pass
 
     def get_placeholder_mask(self, input_ids: torch.LongTensor, inputs_embeds: torch.FloatTensor):
         """

src/transformers/models/internvl/convert_internvl_weights_to_hf.py

@@ -169,11 +169,7 @@ def convert_old_keys_to_new_keys(state_dict_keys: Optional[dict] = None, path: O
                 new_text = re.sub(pattern, replacement, new_text)
         output_dict.update(dict(zip(old_text_language.split("\n"), new_text.split("\n"))))
         old_text_multi = "\n".join(
-            [
-                key
-                for key in state_dict_keys
-                if not (key.startswith("language_model") or key.startswith("vision_model"))
-            ]
+            [key for key in state_dict_keys if not (key.startswith(("language_model", "vision_model")))]
         )
         new_text = old_text_multi
         for pattern, replacement in ORIGINAL_TO_CONVERTED_KEY_MAPPING_MULTI.items():

src/transformers/models/luke/convert_luke_original_pytorch_checkpoint_to_pytorch.py

@@ -75,10 +75,10 @@ def convert_luke_checkpoint(checkpoint_path, metadata_path, entity_vocab_path, p
     missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False)
     if not (len(missing_keys) == 1 and missing_keys[0] == "embeddings.position_ids"):
         raise ValueError(f"Missing keys {', '.join(missing_keys)}. Expected only missing embeddings.position_ids")
-    if not (all(key.startswith("entity_predictions") or key.startswith("lm_head") for key in unexpected_keys)):
+    if not (all(key.startswith(("entity_predictions", "lm_head")) for key in unexpected_keys)):
         raise ValueError(
             "Unexpected keys"
-            f" {', '.join([key for key in unexpected_keys if not (key.startswith('entity_predictions') or key.startswith('lm_head'))])}"
+            f" {', '.join([key for key in unexpected_keys if not (key.startswith(('entity_predictions', 'lm_head')))])}"
         )
 
     # Check outputs

src/transformers/models/marian/convert_marian_to_pytorch.py

@@ -574,11 +574,14 @@ def _check_layer_entries(self):
     def extra_keys(self):
         extra = []
         for k in self.state_keys:
-            if (
-                k.startswith("encoder_l")
-                or k.startswith("decoder_l")
-                or k in [CONFIG_KEY, "Wemb", "encoder_Wemb", "decoder_Wemb", "Wpos", "decoder_ff_logit_out_b"]
-            ):
+            if k.startswith(("encoder_l", "decoder_l")) or k in [
+                CONFIG_KEY,
+                "Wemb",
+                "encoder_Wemb",
+                "decoder_Wemb",
+                "Wpos",
+                "decoder_ff_logit_out_b",
+            ]:
                 continue
             else:
                 extra.append(k)

src/transformers/models/maskformer/convert_maskformer_original_pytorch_checkpoint_to_pytorch.py

@@ -371,7 +371,7 @@ def rename_keys_for_conv(detectron_conv: str, mine_conv: str):
         renamed_keys.extend(rename_keys_for_conv(f"{src_prefix}.layer_4", f"{dst_prefix}.fpn.stem"))
 
         # add all the fpn layers (here we need some config parameters to know the size in advance)
-        for src_i, dst_i in zip(range(3, 0, -1), range(0, 3)):
+        for src_i, dst_i in zip(range(3, 0, -1), range(3)):
             renamed_keys.extend(
                 rename_keys_for_conv(f"{src_prefix}.adapter_{src_i}", f"{dst_prefix}.fpn.layers.{dst_i}.proj")
             )

src/transformers/models/maskformer/convert_maskformer_resnet_to_pytorch.py

@@ -162,7 +162,7 @@ def create_rename_keys(config):
     rename_keys.append(("sem_seg_head.layer_4.weight", "model.pixel_level_module.decoder.fpn.stem.0.weight"))
     rename_keys.append(("sem_seg_head.layer_4.norm.weight", "model.pixel_level_module.decoder.fpn.stem.1.weight"))
     rename_keys.append(("sem_seg_head.layer_4.norm.bias", "model.pixel_level_module.decoder.fpn.stem.1.bias"))
-    for source_index, target_index in zip(range(3, 0, -1), range(0, 3)):
+    for source_index, target_index in zip(range(3, 0, -1), range(3)):
         rename_keys.append((f"sem_seg_head.adapter_{source_index}.weight", f"model.pixel_level_module.decoder.fpn.layers.{target_index}.proj.0.weight"))
         rename_keys.append((f"sem_seg_head.adapter_{source_index}.norm.weight", f"model.pixel_level_module.decoder.fpn.layers.{target_index}.proj.1.weight"))
         rename_keys.append((f"sem_seg_head.adapter_{source_index}.norm.bias", f"model.pixel_level_module.decoder.fpn.layers.{target_index}.proj.1.bias"))

src/transformers/models/maskformer/convert_maskformer_swin_to_pytorch.py

@@ -109,7 +109,7 @@ def create_rename_keys(config):
     rename_keys.append(("sem_seg_head.layer_4.weight", "model.pixel_level_module.decoder.fpn.stem.0.weight"))
     rename_keys.append(("sem_seg_head.layer_4.norm.weight", "model.pixel_level_module.decoder.fpn.stem.1.weight"))
     rename_keys.append(("sem_seg_head.layer_4.norm.bias", "model.pixel_level_module.decoder.fpn.stem.1.bias"))
-    for source_index, target_index in zip(range(3, 0, -1), range(0, 3)):
+    for source_index, target_index in zip(range(3, 0, -1), range(3)):
         rename_keys.append((f"sem_seg_head.adapter_{source_index}.weight", f"model.pixel_level_module.decoder.fpn.layers.{target_index}.proj.0.weight"))
         rename_keys.append((f"sem_seg_head.adapter_{source_index}.norm.weight", f"model.pixel_level_module.decoder.fpn.layers.{target_index}.proj.1.weight"))
         rename_keys.append((f"sem_seg_head.adapter_{source_index}.norm.bias", f"model.pixel_level_module.decoder.fpn.layers.{target_index}.proj.1.bias"))

src/transformers/models/mluke/convert_mluke_original_pytorch_checkpoint_to_pytorch.py

@@ -101,7 +101,7 @@ def convert_luke_checkpoint(checkpoint_path, metadata_path, entity_vocab_path, p
     state_dict.pop("lm_head.decoder.bias")
     state_dict_for_hugging_face = OrderedDict()
     for key in state_dict:
-        if not (key.startswith("lm_head") or key.startswith("entity_predictions")):
+        if not (key.startswith(("lm_head", "entity_predictions"))):
             state_dict_for_hugging_face[f"luke.{key}"] = state_dict[key]
         else:
             state_dict_for_hugging_face[key] = state_dict[key]

src/transformers/models/mm_grounding_dino/convert_mm_grounding_dino_to_hf.py

@@ -377,10 +377,7 @@ def preprocess_old_state(state_dict: dict, config: MMGroundingDinoConfig) -> dic
             k == "dn_query_generator.label_embedding.weight"
             or k == "language_model.language_backbone.body.model.embeddings.position_ids"
             or k == "image_seperate.weight"
-            or k.startswith("lmm")
-            or k.startswith("connector")
-            or k.startswith("region_connector")
-            or k.startswith("ref_point_head")
+            or k.startswith(("lmm", "connector", "region_connector", "ref_point_head"))
         ):
             new_state_dict.pop(k)
 

src/transformers/models/mobilevit/convert_mlcvnets_to_pytorch.py

@@ -95,13 +95,13 @@ def rename_key(name, base_model=False):
     if ".conv_proj." in name:
         name = name.replace(".conv_proj.", ".conv_projection.")
 
-    for i in range(0, 2):
-        for j in range(0, 4):
+    for i in range(2):
+        for j in range(4):
             if f".{i}.{j}." in name:
                 name = name.replace(f".{i}.{j}.", f".{i}.layer.{j}.")
 
     for i in range(2, 6):
-        for j in range(0, 4):
+        for j in range(4):
             if f".{i}.{j}." in name:
                 name = name.replace(f".{i}.{j}.", f".{i}.")
                 if "expand_1x1" in name:

src/transformers/models/mobilevitv2/convert_mlcvnets_to_pytorch.py

@@ -242,7 +242,7 @@ def convert_mobilevitv2_checkpoint(task_name, checkpoint_path, orig_config_path,
     checkpoint = torch.load(checkpoint_path, map_location="cpu", weights_only=True)
 
     # load huggingface model
-    if task_name.startswith("ade20k_") or task_name.startswith("voc_"):
+    if task_name.startswith(("ade20k_", "voc_")):
         model = MobileViTV2ForSemanticSegmentation(config).eval()
         base_model = False
     else:

src/transformers/models/nougat/tokenization_nougat_fast.py

@@ -191,7 +191,7 @@ def truncate_repetitions(text: str, min_len: int = 30) -> str:
     for repetition_length in range(min_len, int(text_length / 2)):
         # check if there is a repetition at the end
         same = True
-        for i in range(0, repetition_length):
+        for i in range(repetition_length):
             if text_lower[text_length - repetition_length - i - 1] != text_lower[text_length - i - 1]:
                 same = False
                 break