support SDPA Attention in stablelm

docs/source/en/perf_infer_gpu_one.md

@@ -176,6 +176,7 @@ For now, Transformers supports SDPA inference and training for the following arc
 * [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel)
 * [Mistral](https://huggingface.co/docs/transformers/model_doc/mistral#transformers.MistralModel)
 * [Mixtral](https://huggingface.co/docs/transformers/model_doc/mixtral#transformers.MixtralModel)
+* [StableLm](https://huggingface.co/docs/transformers/model_doc/stablelm#transformers.StableLmModel)
 * [Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2#transformers.Qwen2Model)
 
 <Tip>

src/transformers/models/stablelm/modeling_stablelm.py

@@ -29,7 +29,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
-from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
 from ...modeling_utils import PreTrainedModel
 from ...utils import (
@@ -374,6 +374,102 @@ def forward(
         return attn_output, attn_weights, past_key_value
 
 
+class StableLmSdpaAttention(StableLmAttention):
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "StableLmModel is using StableLmSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+
+        # Partial rotary embedding
+        query_rot, query_pass = (
+            query_states[..., : self.rotary_emb.dim],
+            query_states[..., self.rotary_emb.dim :],
+        )
+        key_rot, key_pass = (
+            key_states[..., : self.rotary_emb.dim],
+            key_states[..., self.rotary_emb.dim :],
+        )
+        # [batch_size, seq_length, num_heads, head_dim // config.partial_rotary_factor]
+        query_rot, key_rot = apply_rotary_pos_emb(query_rot, key_rot, cos, sin, position_ids)
+
+        # [batch_size, seq_length, num_heads, head_dim]
+        query_states = torch.cat((query_rot, query_pass), dim=-1)
+        key_states = torch.cat((key_rot, key_pass), dim=-1)
+
+        if past_key_value is not None:
+            # Specific to RoPE models with partial rotation
+            cache_kwargs = {"sin": sin, "cos": cos, "partial_rotation_size": self.rotary_emb.dim}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # Repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.attention_dropout.p if self.training else 0.0,
+            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+            is_causal=self.is_causal and attention_mask is None and q_len > 1,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
 class StableLmFlashAttention2(StableLmAttention):
     """
     StableLM flash attention module. This module inherits from `StableLmAttention` as the weights of the module stays
@@ -574,6 +670,7 @@ def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query
 
 ATTENTION_CLASSES = {
     "eager": StableLmAttention,
+    "sdpa": StableLmSdpaAttention,
     "flash_attention_2": StableLmFlashAttention2,
 }
 
@@ -680,6 +777,7 @@ class StableLmPreTrainedModel(PreTrainedModel):
     _skip_keys_device_placement = "past_key_values"
     _supports_flash_attn_2 = True
     _supports_cache_class = True
+    _supports_sdpa = True
 
     def _init_weights(self, module):
         std = self.config.initializer_range
@@ -858,6 +956,11 @@ def forward(
         if self._attn_implementation == "flash_attention_2":
             # 2d mask is passed through the layers
             attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+        # for output_attentions case used fallback to eager attention realization
+        elif self._attn_implementation == "sdpa" and not output_attentions:
+            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+            )
         else:
             # 4d mask is passed through the layers
             attention_mask = _prepare_4d_causal_attention_mask(

tests/models/stablelm/test_modeling_stablelm.py

@@ -24,6 +24,7 @@
     require_bitsandbytes,
     require_flash_attn,
     require_torch,
+    require_torch_sdpa,
     slow,
     torch_device,
 )
@@ -431,3 +432,65 @@ def test_model_3b_long_prompt(self):
         input_ids = torch.tensor([input_ids]).to(model.model.embed_tokens.weight.device)
         generated_ids = model.generate(input_ids, max_new_tokens=4, temperature=0)
         self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-3:].tolist())
+
+    # Copied from transformers.tests.models.llama.test_modeling_llama.LlamaModelTest.test_eager_matches_sdpa_generate with Llama->StableLm,saibo/llama-1B->stabilityai/stablelm-3b-4e1t
+    @require_torch_sdpa
+    @slow
+    def test_eager_matches_sdpa_generate(self):
+        """
+        Overwritting the common test as the test is flaky on tiny models
+        """
+        max_new_tokens = 30
+
+        tokenizer = AutoTokenizer.from_pretrained("stabilityai/stablelm-3b-4e1t")
+
+        model_sdpa = StableLmForCausalLM.from_pretrained(
+            "stabilityai/stablelm-3b-4e1t",
+            torch_dtype=torch.float16,
+            low_cpu_mem_usage=True,
+        ).to(torch_device)
+
+        self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
+
+        model_eager = StableLmForCausalLM.from_pretrained(
+            "stabilityai/stablelm-3b-4e1t",
+            torch_dtype=torch.float16,
+            low_cpu_mem_usage=True,
+            attn_implementation="eager",
+        ).to(torch_device)
+
+        self.assertTrue(model_eager.config._attn_implementation == "eager")
+
+        for name, submodule in model_eager.named_modules():
+            if "SdpaAttention" in submodule.__class__.__name__:
+                raise ValueError("The eager model should not have SDPA attention layers")
+
+        has_sdpa = False
+        for name, submodule in model_sdpa.named_modules():
+            if "SdpaAttention" in submodule.__class__.__name__:
+                has_sdpa = True
+                break
+        if not has_sdpa:
+            raise ValueError("The SDPA model should have SDPA attention layers")
+
+        texts = [
+            "hi here's a longer context, getting longer and",
+            "Hello this is a very long sentence my friend, very long for real",
+            "Today I am in Paris and",
+        ]
+
+        for padding_side in ["left", "right"]:
+            tokenizer.padding_side = padding_side
+            tokenizer.pad_token = tokenizer.eos_token
+
+            inputs = tokenizer(texts, return_tensors="pt", padding=True).to(torch_device)
+
+            res_eager = model_eager.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
+            res_sdpa = model_sdpa.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
+
+            with self.subTest(f"{padding_side}"):
+                torch.testing.assert_close(
+                    res_eager,
+                    res_sdpa,
+                    msg=f"\n{tokenizer.batch_decode(res_eager)} \nvs\n{tokenizer.batch_decode(res_sdpa)}",
+                )