add patching for update_causal_mask to falcon for >= 4.45 (#989)

* add patching for update_causal_mask to falcon and gpt-like models for >=4.45

* fix falcon

* enable codegen2 back

* Apply suggestions from code review

Co-authored-by: Nikita Savelyev <[email protected]>

* Update optimum/exporters/openvino/model_patcher.py

---------

Co-authored-by: Nikita Savelyev <[email protected]>

Author: eaidova

optimum/exporters/openvino/model_configs.py

@@ -29,6 +29,7 @@
     CodeGenOnnxConfig,
     FalconOnnxConfig,
     GemmaOnnxConfig,
+    GPTJOnnxConfig,
     GPTNeoXOnnxConfig,
     IBertOnnxConfig,
     LlamaOnnxConfig,
@@ -66,6 +67,7 @@
     FalconModelPatcher,
     FluxTransfromerModelPatcher,
     Gemma2ModelPatcher,
+    GptJModelPatcher,
     GptNeoxJapaneseModelPatcher,
     GptNeoxModelPatcher,
     IBertModelPatcher,
@@ -726,6 +728,24 @@ def patch_model_for_export(
         return GptNeoxJapaneseModelPatcher(self, model, model_kwargs=model_kwargs)
 
 
+@register_in_tasks_manager(
+    "gptj",
+    *[
+        "feature-extraction",
+        "feature-extraction-with-past",
+        "text-generation",
+        "text-generation-with-past",
+        "text-classification",
+    ],
+    library_name="transformers",
+)
+class GPTJOpenVINOConfig(GPTJOnnxConfig):
+    def patch_model_for_export(
+        self, model: Union["PreTrainedModel", "TFPreTrainedModel"], model_kwargs: Optional[Dict[str, Any]] = None
+    ) -> "ModelPatcher":
+        return GptJModelPatcher(self, model, model_kwargs=model_kwargs)
+
+
 @register_in_tasks_manager(
     "cohere",
     *[

optimum/exporters/openvino/model_patcher.py

@@ -109,11 +109,20 @@ def patch_model_with_bettertransformer(model):
     return model
 
 
-def patch_update_causal_mask(model, transformers_version):
+def patch_update_causal_mask(model, transformers_version, inner_model_name="model", patch_fn=None):
     if is_transformers_version(">=", transformers_version):
-        inner_model = getattr(model, "model", getattr(model, "transformer", None))
+        inner_model = getattr(model, inner_model_name, None)
         if inner_model is not None:
-            inner_model._update_causal_mask = types.MethodType(_llama_gemma_update_causal_mask, inner_model)
+            if hasattr(inner_model, "_update_causal_mask"):
+                inner_model._orig_update_causal_mask = inner_model._update_causal_mask
+            patch_fn = patch_fn or _llama_gemma_update_causal_mask
+            inner_model._update_causal_mask = types.MethodType(patch_fn, inner_model)
+
+
+def unpatch_update_causal_mask(model, inner_model_name="model"):
+    inner_model = getattr(model, inner_model_name, None)
+    if inner_model is not None and hasattr(inner_model, "._orig_update_causal_mask"):
+        inner_model._update_causal_mask = inner_model._orig_update_causal_mask
 
 
 # initialization of sin/cos cached in bf16/fp16 leads to accuracy loss
@@ -579,13 +588,11 @@ def __enter__(self):
 
         # llama/gemma has some accuracy issues with bf16 with transformers >= 4.39
         # fill causal mask in slightly different way for avoid overflow on some platforms
-        patch_update_causal_mask(self._model, "4.39.0")
+        patch_update_causal_mask(self._model, "4.39.0", "model" if hasattr(self._model, "model") else "transformer")
 
     def __exit__(self, exc_type, exc_value, traceback):
         super().__exit__(exc_type, exc_value, traceback)
-        inner_model = getattr(self._model, "model", getattr(self._model, "transformer", None))
-        if hasattr(inner_model, "_orig_update_causal_mask"):
-            inner_model._update_causal_mask = inner_model._orig_update_causal_mask
+        unpatch_update_causal_mask(self._model, "model" if hasattr(self._model, "model") else "transformer")
 
 
 # copied from https://github.com/huggingface/transformers/commit/57d7594a79a9f5d835abf2d4d384db0e4818e548 to unblock export with transformers 4.42
@@ -1865,6 +1872,67 @@ def __exit__(self, exc_type, exc_value, traceback):
                 layer.self_attn.forward = layer.self_attn._orig_forward
 
 
+# copied from  https://github.com/huggingface/optimum/blob/2112e99122d7f23a1da1a9d263fef64301050ea7/optimum/bettertransformer/models/attention.py#L168
+# for preserving backward compatibility between outdated codegen remote code and new transformers
+def _codegen_wrapped_scaled_dot_product_legacy(
+    self,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    head_mask: Optional[torch.Tensor] = None,
+):
+    from optimum.bettertransformer.models.attention import raise_on_head_mask
+
+    raise_on_head_mask(head_mask)
+    batch_size = query.shape[0]
+    mask_value = torch.finfo(value.dtype).min
+    mask_value = torch.full([], mask_value, dtype=value.dtype)
+
+    if batch_size == 1 and attention_mask is not None and attention_mask[0, 0, -1, -1] < -1:
+        raise ValueError("BetterTransformer does not support padding='max_length' with a batch size of 1.")
+
+    # in codegen the query and key are always in fp32 regardless of the dtype of the model
+    # https://github.com/huggingface/transformers/blob/5b28b7833297adf65c5160a685425ddb1eee5ce2/src/transformers/models/codegen/modeling_codegen.py#L226
+    query = query.to(value.dtype)
+    key = key.to(value.dtype)
+
+    dropout_p = self.dropout_prob_attn if self.training else 0.0
+    if batch_size == 1 or self.training:
+        if query.shape[2] > 1:
+            # first step of the decoding
+            sdpa_result = torch.nn.functional.scaled_dot_product_attention(
+                query, key, value, attn_mask=None, dropout_p=dropout_p, is_causal=True
+            )
+        else:
+            # in this case, which is the later decoding steps, the `causal_mask`` in
+            # https://github.com/huggingface/transformers/blob/ae54e3c3b18bac0832ad62ea9b896dfd52a09850/src/transformers/models/gpt2/modeling_gpt2.py#L195
+            # is [True, ..., True] so actually not causal
+            sdpa_result = torch.nn.functional.scaled_dot_product_attention(
+                query, key, value, attn_mask=None, dropout_p=dropout_p, is_causal=False
+            )
+    else:
+        query_length, key_length = query.size(-2), key.size(-2)
+
+        # causal_mask is always [True, ..., True] otherwise, so executing this is unnecessary
+        if query_length > 1:
+            causal_mask = self.causal_mask[:, :, key_length - query_length : key_length, :key_length].to(torch.bool)
+
+            causal_mask = torch.where(causal_mask, 0, mask_value)
+
+            # torch.Tensor.expand does no memory copy
+            causal_mask = causal_mask.expand(batch_size, -1, -1, -1)
+
+            # we use torch.min to avoid having tensor(-inf)
+            attention_mask = torch.min(causal_mask, attention_mask)
+
+        sdpa_result = torch.nn.functional.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=dropout_p, is_causal=False
+        )
+
+    return sdpa_result, None
+
+
 class CodeGenModelPatcher(DecoderModelPatcher):
     def __enter__(self):
         super().__enter__()
@@ -1873,14 +1941,23 @@ def __enter__(self):
         # For avoiding breaking model on tracing stage, we reduce area of bettertransformer patch only for _attn.
         from optimum.bettertransformer.models.attention import codegen_wrapped_scaled_dot_product
 
+        attn_fn = codegen_wrapped_scaled_dot_product
+        if is_torch_version(">=", "2.1.0") and is_transformers_version(">=", "4.45"):
+            # in transformers 4.45 causal_mask const buffer was removed from the model
+            # if it still exists, it means legacy remote code was loaded
+            if hasattr(self._model.transformer.h[0].attn, "causal_mask"):
+                attn_fn = _codegen_wrapped_scaled_dot_product_legacy
+
         for layer in self._model.transformer.h:
             if is_torch_version(">=", "2.1.0") and not self._model.config.output_attentions:
                 orig_self_attn_fwd = layer.attn._attn
-                layer.attn._attn = types.MethodType(codegen_wrapped_scaled_dot_product, layer.attn)
+                layer.attn._attn = types.MethodType(attn_fn, layer.attn)
                 layer.attn._orig_attn = orig_self_attn_fwd
+        patch_update_causal_mask(self._model, "4.45.0", "transformer")
 
     def __exit__(self, exc_type, exc_value, traceback):
         super().__exit__(exc_type, exc_value, traceback)
+        unpatch_update_causal_mask(self._model, "transformer")
         for layer in self._model.transformer.h:
             if hasattr(layer.attn, "_orig_attn"):
                 layer.attn._attn = layer.attn._orig_attn
@@ -2275,8 +2352,7 @@ def __enter__(self):
 
     def __exit__(self, exc_type, exc_value, traceback):
         super().__exit__(exc_type, exc_value, traceback)
-        if hasattr(self._model.model, "_orig_update_causal_mask"):
-            self._model.model._update_causal_mask = self._model.model._orig_update_causal_mask
+        unpatch_update_causal_mask(self._model)
         for layer in self._model.model.layers:
             if hasattr(layer.self_attn, "_orig_forward"):
                 layer.self_attn.forward = layer.self_attn._orig_forward
@@ -2413,8 +2489,7 @@ def __enter__(self):
 
     def __exit__(self, exc_type, exc_value, traceback):
         super().__exit__(exc_type, exc_value, traceback)
-        if hasattr(self._model.model, "_orig_update_causal_mask"):
-            self._model.model._update_causal_mask = self._model.model._orig_update_causal_mask
+        unpatch_update_causal_mask(self._model)
 
 
 class RotaryEmbPatcher(DecoderModelPatcher):
@@ -2425,12 +2500,119 @@ def __enter__(self):
                 _reinitialize_cos_sin_cached_fp32(layer.self_attn.rotary_emb)
 
 
+def _falcon_update_causal_mask(
+    self,
+    attention_mask: torch.Tensor,
+    input_tensor: torch.Tensor,
+    cache_position: torch.Tensor,
+    past_key_values: "Cache",
+    output_attentions: bool,
+    head_mask: torch.Tensor,
+    alibi: torch.Tensor,
+):
+    # copied from  https://github.com/huggingface/transformers/blob/a30c865f991dfec9452cc64bd9a97bfbb96be036/src/transformers/models/falcon/modeling_falcon.py#L1130
+    from transformers.cache_utils import StaticCache
+    from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+
+    # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
+    # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
+    # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+    # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+
+    if hasattr(self, "_prepare_4d_causal_attention_mask_with_cache_position"):
+        _prepare_4d_causal_attention_mask_with_cache_position = (
+            self._prepare_4d_causal_attention_mask_with_cache_position
+        )
+    else:
+        from transformers.models.falcon.modeling_falcon import _prepare_4d_causal_attention_mask_with_cache_position
+
+    if self.config._attn_implementation == "flash_attention_2":
+        if attention_mask is not None and 0.0 in attention_mask:
+            return attention_mask
+        return None
+
+    # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+    # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+    # to infer the attention mask.
+    past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+    using_static_cache = isinstance(past_key_values, StaticCache)
+
+    # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+    if (
+        self.config._attn_implementation == "sdpa"
+        and not using_static_cache
+        and not output_attentions
+        and head_mask is None
+        and alibi is None
+    ):
+        if AttentionMaskConverter._ignore_causal_mask_sdpa(
+            attention_mask,
+            inputs_embeds=input_tensor,
+            past_key_values_length=past_seen_tokens,
+            is_training=self.training,
+        ):
+            return None
+
+    dtype, device = input_tensor.dtype, input_tensor.device
+    # difference from original, replace torch.finfo(dtype).min to float16 for prevent overflow for fp16/bf16 execution
+    min_dtype = torch.finfo(torch.float16).min
+    batch_size, sequence_length, _ = input_tensor.shape
+    if using_static_cache:
+        target_length = past_key_values.get_max_length()
+    else:
+        target_length = (
+            attention_mask.shape[-1]
+            if isinstance(attention_mask, torch.Tensor)
+            else past_seen_tokens + sequence_length
+        )
+
+    # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+    causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask,
+        sequence_length=sequence_length,
+        target_length=target_length,
+        dtype=dtype,
+        device=device,
+        min_dtype=min_dtype,
+        cache_position=cache_position,
+        batch_size=input_tensor.shape[0],
+    )
+
+    # We take care to integrate alibi bias in the causal_mask here
+    if head_mask is None and alibi is not None:
+        alibi = alibi.reshape(batch_size, -1, *alibi.shape[1:])
+        causal_mask = torch.masked_fill(
+            alibi / math.sqrt(self.config.hidden_size // self.num_heads),
+            causal_mask < -1,
+            min_dtype,
+        )
+
+    if (
+        self.config._attn_implementation == "sdpa"
+        and attention_mask is not None
+        and attention_mask.device.type == "cuda"
+        and not output_attentions
+    ):
+        # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+        # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+        # Details: https://github.com/pytorch/pytorch/issues/110213
+        causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+    return causal_mask
+
+
 class FalconModelPatcher(DecoderModelPatcher):
     def __enter__(self):
         super().__enter__()
         if is_transformers_version("<", "4.44.99"):
             for layer in self._model.transformer.h:
                 _reinitialize_cos_sin_cached_fp32(layer.self_attention.rotary_emb)
+        else:
+            patch_update_causal_mask(self._model, "4.45.0", "transformer", _falcon_update_causal_mask)
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        super().__exit__(exc_type, exc_value, traceback)
+        unpatch_update_causal_mask(self._model, "transformer")
 
 
 class GptNeoxModelPatcher(DecoderModelPatcher):
@@ -2439,6 +2621,22 @@ def __enter__(self):
         if is_transformers_version("<", "4.44.99"):
             for layer in self._model.gpt_neox.layers:
                 _reinitialize_cos_sin_cached_fp32(layer.attention.rotary_emb)
+        else:
+            patch_update_causal_mask(self._model, "4.45.0", "gpt_neox")
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        super().__exit__(exc_type, exc_value, traceback)
+        unpatch_update_causal_mask(self._model, "gpt_neox")
+
+
+class GptJModelPatcher(DecoderModelPatcher):
+    def __enter__(self):
+        super().__enter__()
+        patch_update_causal_mask(self._model, "4.45.0", "transformer")
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        super().__exit__(exc_type, exc_value, traceback)
+        unpatch_update_causal_mask(self._model, "transformer")
 
 
 class GptNeoxJapaneseModelPatcher(DecoderModelPatcher):
@@ -2447,6 +2645,12 @@ def __enter__(self):
         if is_transformers_version("<", "4.44.99"):
             for layer in self._model.gpt_neox_japanese.layers:
                 _reinitialize_cos_sin_cached_fp32(layer.attention.rotary_emb)
+        else:
+            patch_update_causal_mask(self._model, "4.45.0", "gpt_neox_japanese")
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        super().__exit__(exc_type, exc_value, traceback)
+        unpatch_update_causal_mask(self._model, "gpt_neox_japanese")
 
 
 class Gemma2ModelPatcher(LlamaModelPatcher):

tests/openvino/test_modeling.py

@@ -773,6 +773,7 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
         "bloom",
         "chatglm",
         "codegen",
+        "codegen2",
         "gpt2",
         "gpt_neo",
         "gpt_neox",
@@ -821,10 +822,6 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
             "mistral-nemo",
         )
 
-    # custom modeling defined in https://huggingface.co/katuni4ka/tiny-random-codegen2 differs from transformers after v4.45 resulting in unadapted patching
-    if is_transformers_version("<", "4.45.0"):
-        SUPPORTED_ARCHITECTURES += ("codegen2",)
-
     GENERATION_LENGTH = 100
     REMOTE_CODE_MODELS = (
         "chatglm",