transformers 4.43.0__tar.gz → 4.43.2__tar.gz

{transformers-4.43.0 → transformers-4.43.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: transformers
-Version: 4.43.0
+Version: 4.43.2
 Summary: State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow
 Home-page: https://github.com/huggingface/transformers
 Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)

{transformers-4.43.0 → transformers-4.43.2}/setup.py

@@ -430,7 +430,7 @@ install_requires = [
 
 setup(
     name="transformers",
-    version="4.43.0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="4.43.2",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     author="The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)",
     author_email="transformers@huggingface.co",
     description="State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow",

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/__init__.py

@@ -18,7 +18,7 @@
 # to defer the actual importing for when the objects are requested. This way `import transformers` provides the names
 # in the namespace without actually importing anything (and especially none of the backends).
 
-__version__ = "4.43.0"
+__version__ = "4.43.2"
 
 from typing import TYPE_CHECKING
 

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/modeling_rope_utils.py

@@ -354,6 +354,11 @@ ROPE_INIT_FUNCTIONS = {
 
 def _check_received_keys(rope_type: str, received_keys: set, required_keys: set, optional_keys: Optional[set] = None):
     """Compare the received keys in `config.rope_scaling` against the expected and optional keys"""
+    # BC: "rope_type" was originally "type" -- let's gracefully handle it
+    if "rope_type" not in received_keys and "type" in received_keys:
+        received_keys -= {"type"}
+        received_keys.add("rope_type")
+
     missing_keys = required_keys - received_keys
     if missing_keys:
         raise KeyError(f"Missing required keys in `rope_scaling` for 'rope_type'='{rope_type}': {missing_keys}")
@@ -361,14 +366,14 @@ def _check_received_keys(rope_type: str, received_keys: set, required_keys: set,
     if optional_keys is not None:
         unused_keys = received_keys - required_keys - optional_keys
     else:
-        unused_keys = received_keys - received_keys
+        unused_keys = received_keys - required_keys
     if unused_keys:
-        raise KeyError(f"Unrecognized keys in `rope_scaling` for 'rope_type'='{rope_type}': {unused_keys}")
+        logger.warning(f"Unrecognized keys in `rope_scaling` for 'rope_type'='{rope_type}': {unused_keys}")
 
 
 def _validate_default_rope_parameters(config: PretrainedConfig):
     rope_scaling = config.rope_scaling
-    rope_type = rope_scaling["rope_type"]
+    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))  # BC: "rope_type" was originally "type"
     required_keys = {"rope_type"}
     received_keys = set(rope_scaling.keys())
     _check_received_keys(rope_type, received_keys, required_keys)
@@ -376,19 +381,19 @@ def _validate_default_rope_parameters(config: PretrainedConfig):
 
 def _validate_linear_scaling_rope_parameters(config: PretrainedConfig):
     rope_scaling = config.rope_scaling
-    rope_type = rope_scaling["rope_type"]
+    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))  # BC: "rope_type" was originally "type"
     required_keys = {"rope_type", "factor"}
     received_keys = set(rope_scaling.keys())
     _check_received_keys(rope_type, received_keys, required_keys)
 
     factor = rope_scaling["factor"]
     if factor is None or not isinstance(factor, float) or factor < 1.0:
-        raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
+        logger.warning(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
 
 
 def _validate_dynamic_scaling_rope_parameters(config: PretrainedConfig):
     rope_scaling = config.rope_scaling
-    rope_type = rope_scaling["rope_type"]
+    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))  # BC: "rope_type" was originally "type"
     required_keys = {"rope_type", "factor"}
     # TODO (joao): update logic for the inclusion of `original_max_position_embeddings`
     optional_keys = {"original_max_position_embeddings"}
@@ -397,12 +402,12 @@ def _validate_dynamic_scaling_rope_parameters(config: PretrainedConfig):
 
     factor = rope_scaling["factor"]
     if factor is None or not isinstance(factor, float) or factor < 1.0:
-        raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
+        logger.warning(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
 
 
 def _validate_yarn_parameters(config: PretrainedConfig):
     rope_scaling = config.rope_scaling
-    rope_type = rope_scaling["rope_type"]
+    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))  # BC: "rope_type" was originally "type"
     required_keys = {"rope_type", "factor"}
     optional_keys = {"attention_factor", "beta_fast", "beta_slow"}
     received_keys = set(rope_scaling.keys())
@@ -410,22 +415,22 @@ def _validate_yarn_parameters(config: PretrainedConfig):
 
     factor = rope_scaling["factor"]
     if factor is None or not isinstance(factor, float) or factor < 1.0:
-        raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
+        logger.warning(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
 
     attention_factor = rope_scaling.get("attention_factor")
     if attention_factor is not None and (not isinstance(attention_factor, float) or attention_factor < 0):
-        raise ValueError(
+        logger.warning(
             f"`rope_scaling`'s attention_factor field must be a float greater than 0, got {attention_factor}"
         )
     beta_fast = rope_scaling.get("beta_fast")
     if beta_fast is not None and not isinstance(beta_fast, float):
-        raise ValueError(f"`rope_scaling`'s beta_fast field must be a float, got {beta_fast}")
+        logger.warning(f"`rope_scaling`'s beta_fast field must be a float, got {beta_fast}")
     beta_slow = rope_scaling.get("beta_slow")
     if beta_slow is not None and not isinstance(beta_slow, float):
-        raise ValueError(f"`rope_scaling`'s beta_slow field must be a float, got {beta_slow}")
+        logger.warning(f"`rope_scaling`'s beta_slow field must be a float, got {beta_slow}")
 
     if (beta_fast or 32) < (beta_slow or 1):
-        raise ValueError(
+        logger.warning(
             f"`rope_scaling`'s beta_fast field must be greater than beta_slow, got beta_fast={beta_fast} "
             f"(defaults to 32 if None) and beta_slow={beta_slow} (defaults to 1 if None)"
         )
@@ -433,7 +438,7 @@ def _validate_yarn_parameters(config: PretrainedConfig):
 
 def _validate_longrope_parameters(config: PretrainedConfig):
     rope_scaling = config.rope_scaling
-    rope_type = rope_scaling["rope_type"]
+    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))  # BC: "rope_type" was originally "type"
     required_keys = {"rope_type", "short_factor", "long_factor"}
     # TODO (joao): update logic for the inclusion of `original_max_position_embeddings`
     optional_keys = {"attention_factor", "factor", "original_max_position_embeddings"}
@@ -445,15 +450,15 @@ def _validate_longrope_parameters(config: PretrainedConfig):
 
     short_factor = rope_scaling.get("short_factor")
     if not isinstance(short_factor, list) and all(isinstance(x, (int, float)) for x in short_factor):
-        raise ValueError(f"`rope_scaling`'s short_factor field must be a list of numbers, got {short_factor}")
+        logger.warning(f"`rope_scaling`'s short_factor field must be a list of numbers, got {short_factor}")
     if not len(short_factor) == dim // 2:
-        raise ValueError(f"`rope_scaling`'s short_factor field must have length {dim // 2}, got {len(short_factor)}")
+        logger.warning(f"`rope_scaling`'s short_factor field must have length {dim // 2}, got {len(short_factor)}")
 
     long_factor = rope_scaling.get("long_factor")
     if not isinstance(long_factor, list) and all(isinstance(x, (int, float)) for x in long_factor):
-        raise ValueError(f"`rope_scaling`'s long_factor field must be a list of numbers, got {long_factor}")
+        logger.warning(f"`rope_scaling`'s long_factor field must be a list of numbers, got {long_factor}")
     if not len(long_factor) == dim // 2:
-        raise ValueError(f"`rope_scaling`'s long_factor field must have length {dim // 2}, got {len(long_factor)}")
+        logger.warning(f"`rope_scaling`'s long_factor field must have length {dim // 2}, got {len(long_factor)}")
 
     # Handle Phi3 divergence: prefer the use of `attention_factor` and/or `factor` over
     # `original_max_position_embeddings` to compute internal variables. The latter lives outside `rope_scaling` and is
@@ -468,48 +473,48 @@ def _validate_longrope_parameters(config: PretrainedConfig):
     else:
         factor = rope_scaling.get("factor")
         if factor is None:
-            raise ValueError("Missing required keys in `rope_scaling`: 'factor'")
+            logger.warning("Missing required keys in `rope_scaling`: 'factor'")
         elif not isinstance(factor, float) or factor < 1.0:
-            raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
+            logger.warning(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
 
         attention_factor = rope_scaling.get("attention_factor")
         if attention_factor is not None and not isinstance(attention_factor, float) or attention_factor < 0:
-            raise ValueError(
+            logger.warning(
                 f"`rope_scaling`'s attention_factor field must be a float greater than 0, got {attention_factor}"
             )
 
 
 def _validate_llama3_parameters(config: PretrainedConfig):
     rope_scaling = config.rope_scaling
-    rope_type = rope_scaling["rope_type"]
+    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))  # BC: "rope_type" was originally "type"
     required_keys = {"rope_type", "factor", "original_max_position_embeddings", "low_freq_factor", "high_freq_factor"}
     received_keys = set(rope_scaling.keys())
     _check_received_keys(rope_type, received_keys, required_keys)
 
     factor = rope_scaling["factor"]
     if factor is None or not isinstance(factor, float) or factor < 1.0:
-        raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
+        logger.warning(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
 
     low_freq_factor = rope_scaling["low_freq_factor"]
     high_freq_factor = rope_scaling["high_freq_factor"]
     if low_freq_factor is None or not isinstance(low_freq_factor, float):
-        raise ValueError(f"`rope_scaling`'s low_freq_factor field must be a float, got {low_freq_factor}")
+        logger.warning(f"`rope_scaling`'s low_freq_factor field must be a float, got {low_freq_factor}")
     if high_freq_factor is None or not isinstance(high_freq_factor, float):
-        raise ValueError(f"`rope_scaling`'s high_freq_factor field must be a float, got {high_freq_factor}")
+        logger.warning(f"`rope_scaling`'s high_freq_factor field must be a float, got {high_freq_factor}")
     if high_freq_factor < low_freq_factor:
-        raise ValueError(
+        logger.warning(
             "`rope_scaling`'s high_freq_factor field must be greater than low_freq_factor, got high_freq_factor="
             f"{high_freq_factor} and low_freq_factor={low_freq_factor}"
         )
 
     original_max_position_embeddings = rope_scaling["original_max_position_embeddings"]
     if original_max_position_embeddings is None or not isinstance(original_max_position_embeddings, int):
-        raise ValueError(
+        logger.warning(
             "`rope_scaling`'s original_max_position_embeddings field must be an integer, got "
             f"{original_max_position_embeddings}"
         )
     if original_max_position_embeddings >= config.max_position_embeddings:
-        raise ValueError(
+        logger.warning(
             "`rope_scaling`'s original_max_position_embeddings field must be less than max_position_embeddings, got "
             f"{original_max_position_embeddings} and max_position_embeddings={config.max_position_embeddings}"
         )
@@ -534,17 +539,12 @@ def rope_config_validation(config: PretrainedConfig):
     if rope_scaling is None:
         return
 
-    possible_rope_types = set(ROPE_INIT_FUNCTIONS.keys())
-    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))  # BC: "rope_type" was originally "type"
-    if rope_type is None:
-        raise ValueError(
-            f"rope_scaling must contain a non-None 'rope_type' field. Possible options are {possible_rope_types}"
-        )
-
+    # BC: "rope_type" was originally "type"
+    rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", "default"))
     validation_fn = ROPE_VALIDATION_FUNCTIONS.get(rope_type)
     if validation_fn is not None:
         validation_fn(config)
     else:
-        raise ValueError(
+        logger.warning(
             f"Missing validation function mapping in `ROPE_VALIDATION_FUNCTIONS` for 'rope_type'='{rope_type}'"
         )

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/modeling_utils.py

@@ -855,6 +855,8 @@ def _load_state_dict_into_meta_model(
     for old_key, new_key in zip(old_keys, new_keys):
         state_dict[new_key] = state_dict.pop(old_key)
 
+    is_torch_e4m3fn_available = hasattr(torch, "float8_e4m3fn")
+
     for param_name, param in state_dict.items():
         # First part of the test is always true as load_state_dict_keys always contains state_dict keys.
         if param_name not in loaded_state_dict_keys or param_name not in expected_keys:
@@ -866,9 +868,10 @@ def _load_state_dict_into_meta_model(
         module_name = param_name
         set_module_kwargs = {}
 
-        # We convert floating dtypes to the `dtype` passed. We want to keep the buffers/params
+        # We convert floating dtypes to the `dtype` passed except for float8_e4m3fn type. We also want to keep the buffers/params
         # in int/uint/bool and not cast them.
-        if dtype is not None and torch.is_floating_point(param) and param.dtype != torch.float8_e4m3fn:
+        is_param_float8_e4m3fn = is_torch_e4m3fn_available and param.dtype == torch.float8_e4m3fn
+        if dtype is not None and torch.is_floating_point(param) and not is_param_float8_e4m3fn:
             if (
                 keep_in_fp32_modules is not None
                 and any(
@@ -2131,13 +2134,23 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         # Replace weights in old_embeddings and return to maintain the same embedding type.
         # This ensures correct functionality when a Custom Embedding class is passed as input.
         # The input and output embedding types remain consistent. (c.f. https://github.com/huggingface/transformers/pull/31979)
-        old_embeddings.weight.data = new_embeddings.weight.data
-        old_embeddings.num_embeddings = new_embeddings.weight.data.shape[0]
+        if is_deepspeed_zero3_enabled() and not is_quantized:
+            import deepspeed
 
-        # If the new number of tokens is smaller than the original `padding_idx`, the `padding_idx`
-        # will be set to `None` in the resized embeddings.
-        if old_embeddings.padding_idx is not None and (new_num_tokens - 1) < old_embeddings.padding_idx:
-            old_embeddings.padding_idx = None
+            params = [old_embeddings.weight, new_embeddings.weight]
+            with deepspeed.zero.GatheredParameters(params, modifier_rank=0):
+                old_embeddings.weight.data = new_embeddings.weight.data
+                old_embeddings.num_embeddings = new_embeddings.weight.data.shape[0]
+
+                # If the new number of tokens is smaller than the original `padding_idx`, the `padding_idx`
+                # will be set to `None` in the resized embeddings.
+                if old_embeddings.padding_idx is not None and (new_num_tokens - 1) < old_embeddings.padding_idx:
+                    old_embeddings.padding_idx = None
+        else:
+            old_embeddings.weight.data = new_embeddings.weight.data
+            old_embeddings.num_embeddings = new_embeddings.weight.data.shape[0]
+            if old_embeddings.padding_idx is not None and (new_num_tokens - 1) < old_embeddings.padding_idx:
+                old_embeddings.padding_idx = None
 
         return old_embeddings
 

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/cohere/modeling_cohere.py

@@ -769,7 +769,9 @@ class CohereModel(CoherePreTrainedModel):
 
         past_seen_tokens = 0
         return_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):  # kept for BC (non `Cache` `past_key_values` inputs)
+        if (
+            use_cache and not isinstance(past_key_values, Cache) and not self.training
+        ):  # kept for BC (non `Cache` `past_key_values` inputs)
             return_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
 

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/dbrx/modeling_dbrx.py

@@ -1004,7 +1004,9 @@ class DbrxModel(DbrxPreTrainedModel):
         inputs_embeds = nn.functional.dropout(inputs_embeds, p=self.emb_pdrop, training=self.training)
 
         return_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):  # kept for BC (non `Cache` `past_key_values` inputs)
+        if (
+            use_cache and not isinstance(past_key_values, Cache) and not self.training
+        ):  # kept for BC (non `Cache` `past_key_values` inputs)
             return_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/gemma/diff_gemma.py

@@ -474,7 +474,9 @@ class GemmaModel(LlamaModel):
             inputs_embeds = self.embed_tokens(input_ids)
 
         return_legacy_cache = False  # noqa: F841
-        if use_cache and not isinstance(past_key_values, Cache):  # kept for BC (non `Cache` `past_key_values` inputs)
+        if (
+            use_cache and not isinstance(past_key_values, Cache) and not self.training
+        ):  # kept for BC (non `Cache` `past_key_values` inputs)
             return_legacy_cache = True  # noqa: F841
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
 

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/gemma/modeling_gemma.py

@@ -769,7 +769,9 @@ class GemmaModel(GemmaPreTrainedModel):
             inputs_embeds = self.embed_tokens(input_ids)
 
         return_legacy_cache = False  # noqa: F841
-        if use_cache and not isinstance(past_key_values, Cache):  # kept for BC (non `Cache` `past_key_values` inputs)
+        if (
+            use_cache and not isinstance(past_key_values, Cache) and not self.training
+        ):  # kept for BC (non `Cache` `past_key_values` inputs)
             return_legacy_cache = True  # noqa: F841
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
 
@@ -794,7 +796,9 @@ class GemmaModel(GemmaPreTrainedModel):
         # See https://github.com/huggingface/transformers/pull/29402
         normalizer = torch.tensor(self.config.hidden_size**0.5, dtype=hidden_states.dtype)
         hidden_states = hidden_states * normalizer
-        if use_cache and not isinstance(past_key_values, Cache):  # kept for BC (non `Cache` `past_key_values` inputs)
+        if (
+            use_cache and not isinstance(past_key_values, Cache) and not self.training
+        ):  # kept for BC (non `Cache` `past_key_values` inputs)
             return_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/jetmoe/modeling_jetmoe.py

@@ -978,7 +978,9 @@ class JetMoeModel(JetMoePreTrainedModel):
             inputs_embeds = self.embed_tokens(input_ids)
 
         return_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):  # kept for BC (non `Cache` `past_key_values` inputs)
+        if (
+            use_cache and not isinstance(past_key_values, Cache) and not self.training
+        ):  # kept for BC (non `Cache` `past_key_values` inputs)
             return_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
 

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/llama/configuration_llama.py

@@ -189,6 +189,9 @@ class LlamaConfig(PretrainedConfig):
         self.mlp_bias = mlp_bias
 
         # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
         rope_config_validation(self)
 
         super().__init__(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/llama/modeling_llama.py

@@ -107,7 +107,7 @@ class LlamaRotaryEmbedding(nn.Module):
         else:
             # BC: "rope_type" was originally "type"
             if config.rope_scaling is not None:
-                self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling["type"])
+                self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
             else:
                 self.rope_type = "default"
             self.max_seq_len_cached = config.max_position_embeddings
@@ -893,7 +893,9 @@ class LlamaModel(LlamaPreTrainedModel):
             inputs_embeds = self.embed_tokens(input_ids)
 
         return_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):  # kept for BC (non `Cache` `past_key_values` inputs)
+        if (
+            use_cache and not isinstance(past_key_values, Cache) and not self.training
+        ):  # kept for BC (non `Cache` `past_key_values` inputs)
             return_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/mistral/modeling_mistral.py

@@ -757,7 +757,7 @@ class MistralModel(MistralPreTrainedModel):
             inputs_embeds = self.embed_tokens(input_ids)
 
         return_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
+        if use_cache and not isinstance(past_key_values, Cache) and not self.training:
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             return_legacy_cache = True
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/mixtral/modeling_mixtral.py

@@ -959,7 +959,7 @@ class MixtralModel(MixtralPreTrainedModel):
                 use_cache = False
 
         use_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
+        if use_cache and not isinstance(past_key_values, Cache) and not self.training:
             use_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/olmo/modeling_olmo.py

@@ -810,7 +810,9 @@ class OlmoModel(OlmoPreTrainedModel):
             inputs_embeds = self.embed_tokens(input_ids)
 
         return_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):  # kept for BC (non `Cache` `past_key_values` inputs)
+        if (
+            use_cache and not isinstance(past_key_values, Cache) and not self.training
+        ):  # kept for BC (non `Cache` `past_key_values` inputs)
             return_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/persimmon/modeling_persimmon.py

@@ -626,7 +626,7 @@ class PersimmonModel(PersimmonPreTrainedModel):
                 use_cache = False
 
         use_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
+        if use_cache and not isinstance(past_key_values, Cache) and not self.training:
             use_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/phi/modeling_phi.py

@@ -908,7 +908,7 @@ class PhiModel(PhiPreTrainedModel):
                 use_cache = False
 
         use_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
+        if use_cache and not isinstance(past_key_values, Cache) and not self.training:
             use_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/phi3/modeling_phi3.py

@@ -949,7 +949,7 @@ class Phi3Model(Phi3PreTrainedModel):
                 use_cache = False
 
         use_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
+        if use_cache and not isinstance(past_key_values, Cache) and not self.training:
             use_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/qwen2/modeling_qwen2.py

@@ -807,7 +807,7 @@ class Qwen2Model(Qwen2PreTrainedModel):
                 use_cache = False
 
         use_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
+        if use_cache and not isinstance(past_key_values, Cache) and not self.training:
             use_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py

@@ -969,7 +969,7 @@ class Qwen2MoeModel(Qwen2MoePreTrainedModel):
                 use_cache = False
 
         use_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
+        if use_cache and not isinstance(past_key_values, Cache) and not self.training:
             use_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/stablelm/modeling_stablelm.py

@@ -901,7 +901,7 @@ class StableLmModel(StableLmPreTrainedModel):
                 use_cache = False
 
         use_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
+        if use_cache and not isinstance(past_key_values, Cache) and not self.training:
             use_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers/models/starcoder2/modeling_starcoder2.py

@@ -783,7 +783,7 @@ class Starcoder2Model(Starcoder2PreTrainedModel):
                 use_cache = False
 
         use_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
+        if use_cache and not isinstance(past_key_values, Cache) and not self.training:
             use_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             logger.warning_once(

{transformers-4.43.0 → transformers-4.43.2}/src/transformers.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: transformers
-Version: 4.43.0
+Version: 4.43.2
 Summary: State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow
 Home-page: https://github.com/huggingface/transformers
 Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)