gpt_bigcode: added FusedSDPA kernel

- Added support of following options to gpt_bigcode (starcoder class of models) use_flash_attention, flash_attention_recompute, flash_attention_fast_softmax, flash_attention_causal_mask - Updated test for starcoder model
huggingface · Jul 29, 2024 · 7aa7a8e · 7aa7a8e
1 parent 2cd2e68
commit 7aa7a8e
Show file tree

Hide file tree

Showing 2 changed files with 201 additions and 11 deletions.
diff --git a/optimum/habana/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py b/optimum/habana/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
@@ -1,6 +1,8 @@
+import math
 from typing import List, Optional, Tuple, Union
 
 import torch
+import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions
@@ -9,6 +11,141 @@
 from ...modeling_attn_mask_utils import GaudiAttentionMaskConverter
 
 
+try:
+    from habana_frameworks.torch.hpex.kernels import FusedSDPA
+except ImportError:
+    print("Not using HPU fused scaled dot-product attention kernel.")
+    FusedSDPA = None
+
+import habana_frameworks.torch.core as htcore
+
+
+def gaudi_flash_attn_v1(
+    query_layer, key_layer, value_layer, attention_mask, dropout_rate, is_causal, scale, softmax_mode, q_block_size
+):
+    """
+    Gaudi version of Flash Attention V1 to support long sequence at prompt phase
+    Causal mask is not supported in this optimization
+    """
+    if is_causal:
+        raise ValueError("Causal mask is not supported for long input sequences")
+
+    q_len = query_layer.size(-2)
+    q_tiles = (q_len // q_block_size) if (q_len % q_block_size == 0) else math.ceil(q_len / q_block_size)
+    q_padding = q_tiles * q_block_size - q_len
+    query_layer = F.pad(query_layer, (0, 0, 0, q_padding), "constant", 0)
+    if attention_mask is not None:
+        attention_mask = F.pad(attention_mask, (0, 0, 0, q_padding), "constant", -10000.0)
+    row_o_list = []
+    for i in range(q_tiles):
+        s, e = i * q_block_size, (i + 1) * q_block_size
+        row_q = query_layer[:, :, s:e, :]
+        row_mask = attention_mask[:, :, s:e, :]
+        attn_output_partial = FusedSDPA.apply(
+            row_q, key_layer, value_layer, row_mask, dropout_rate, is_causal, scale, softmax_mode
+        )
+        row_o_list.append(attn_output_partial)
+    attn_output = torch.cat(row_o_list, dim=-2)
+    if q_padding != 0:
+        attn_output = attn_output[:, :, :-q_padding, :]
+    return attn_output
+
+
+def apply_FusedSDPA(
+    self,
+    query,
+    key,
+    value,
+    attention_mask=None,
+    flash_attention_recompute=False,
+    flash_attention_fast_softmax=False,
+    flash_attention_causal_mask=False,
+):
+    """
+    Copied from GPTBigCodeSdpaAttention._attn: https://github.com/huggingface/transformers/blob/v4.40-release/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
+    The only differences are:
+    - replaced torch.nn.functional.scaled_dot_product_attention with Habana's FusedSDPA
+    - removed WA for key and value tensor expanding over heads dimension. That WA also works but dramatically drops throughput
+    - added args use_flash_attention, flash_attention_recompute, flash_attention_fast_softmax, flash_attention_causal_mask to control parameters of FusedSDPA
+    - added special case handling for input larger 8192 with function gaudi_flash_attn_v1
+    """
+
+    scale = None
+    if not self.scale_attn_weights:
+        scale = 1
+
+    # MQA models: (batch_size, query_length, num_heads * head_dim)
+    # MHA models: (batch_size, num_heads, query_length, head_dim)
+    query_shape = query.shape
+    batch_size = query_shape[0]
+
+    if self.multi_query:
+        query_length = query_shape[1]
+
+        # SDPA requires the dimension [..., sequence_length, head_dim].
+        query = query.view(batch_size, query_length, self.num_heads, self.head_dim).transpose(1, 2)
+
+        # Without these unsqueeze, SDPA complains as the query and key/value have a different number of dimensions.
+        key = key.unsqueeze(1)
+        value = value.unsqueeze(1)
+
+    else:
+        query_length = query_shape[-1]
+
+        if attention_mask is not None:
+            query = query.contiguous()
+            key = key.contiguous()
+            value = value.contiguous()
+
+    sdpa_result = None
+    enable_recompute = flash_attention_recompute and query_length > 1
+
+    if query_length > 1 and flash_attention_causal_mask:
+        attention_mask = None
+        use_causal_mask = True
+    else:
+        use_causal_mask = self.is_causal and attention_mask is None and query_length > 1
+
+    import habana_frameworks.torch.hpu as ht
+
+    with ht.sdp_kernel(enable_recompute=enable_recompute):
+        if query_length > 8192:
+            sdpa_result = gaudi_flash_attn_v1(
+                query,
+                key,
+                value,
+                attention_mask,
+                self.attn_pdrop if self.training else 0.0,
+                use_causal_mask,
+                scale,
+                "fast" if flash_attention_fast_softmax else "None",
+                4096,
+            )
+            htcore.mark_step()
+        else:
+            sdpa_result = FusedSDPA.apply(
+                query,
+                key,
+                value,
+                attention_mask,
+                self.attn_pdrop if self.training else 0.0,
+                use_causal_mask,
+                scale,
+                "fast" if flash_attention_fast_softmax else "None",
+            )
+
+    if self.multi_query:
+        # (batch_size, num_heads, seq_len, head_dim) --> (batch_size, seq_len, num_heads, head_dim)
+        sdpa_result = sdpa_result.transpose(1, 2)
+
+        # Reshape is kind of expensive here, as it does a memory copy,
+        # but I did not manage to make away without it (logits do not match when using view)
+        # (batch_size, seq_len, num_heads, head_dim) --> (batch_size, seq_len, num_heads * head_dim)
+        sdpa_result = sdpa_result.reshape(query_shape)
+
+    return sdpa_result, None
+
+
 def gaudi_gpt_bigcode_attention_forward(
     self,
     hidden_states: torch.Tensor,
@@ -20,14 +157,18 @@ def gaudi_gpt_bigcode_attention_forward(
     use_cache: Optional[bool] = False,
     output_attentions: Optional[bool] = False,
     token_idx: Optional[torch.Tensor] = None,
+    use_flash_attention: Optional[bool] = False,
+    flash_attention_recompute: Optional[bool] = False,
+    flash_attention_fast_softmax: Optional[bool] = False,
+    flash_attention_causal_mask: Optional[bool] = False,
 ) -> Union[
     Tuple[torch.Tensor, Optional[torch.Tensor]],
     Tuple[torch.Tensor, Optional[torch.Tensor], Tuple[torch.Tensor, ...]],
 ]:
     """
-    Copied from GPTBigCodeAttention.forward: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
+    Copied from GPTBigCodeAttention.forward: https://github.com/huggingface/transformers/blob/v4.40-release/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
     The only differences are:
-    - add new args token_idx
+    - add new args token_idx, use_flash_attention, flash_attention_recompute, flash_attention_fast_softmax, flash_attention_causal_mask
     - optimize KV cache
     """
     if encoder_hidden_states is not None:
@@ -65,7 +206,21 @@ def gaudi_gpt_bigcode_attention_forward(
             value = torch.cat((past_value, value), dim=-2)
     present = torch.cat((key, value), dim=-1) if use_cache else None
 
-    attn_output, attn_weights = self._attn(query, key.transpose(-1, -2), value, attention_mask, head_mask)
+    if not output_attentions and head_mask is None and use_flash_attention:
+        # Difference with the original implementation: there is no need to transpose the key here,
+        # as SDPA expects seq_length to be at index -2 for the key as well
+        attn_output, attn_weights = apply_FusedSDPA(
+            self,
+            query,
+            key,
+            value,
+            attention_mask,
+            flash_attention_recompute,
+            flash_attention_fast_softmax,
+            flash_attention_causal_mask,
+        )
+    else:
+        attn_output, attn_weights = self._attn(query, key.transpose(-1, -2), value, attention_mask, head_mask)
 
     if not self.multi_query:
         attn_output = attn_output.transpose(1, 2).reshape(hidden_states.shape)
@@ -93,11 +248,15 @@ def gaudi_gpt_bigcode_block_forward(
     use_cache: Optional[bool] = False,
     output_attentions: Optional[bool] = False,
     token_idx: Optional[torch.Tensor] = None,
+    use_flash_attention: Optional[bool] = False,
+    flash_attention_recompute: Optional[bool] = False,
+    flash_attention_fast_softmax: Optional[bool] = False,
+    flash_attention_causal_mask: Optional[bool] = False,
 ) -> Union[Tuple[torch.Tensor], Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor, torch.Tensor]]:
     """
-    Copied from GPTBigCodeBlock.forward: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
+    Copied from GPTBigCodeBlock.forward: https://github.com/huggingface/transformers/blob/v4.40-release/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
     The only differences are:
-    - add new args token_idx
+    - add new args token_idx, use_flash_attention, flash_attention_recompute, flash_attention_fast_softmax, flash_attention_causal_mask
     """
     residual = hidden_states
     hidden_states = self.ln_1(hidden_states)
@@ -109,6 +268,10 @@ def gaudi_gpt_bigcode_block_forward(
         use_cache=use_cache,
         output_attentions=output_attentions,
         token_idx=token_idx,
+        use_flash_attention=use_flash_attention,
+        flash_attention_recompute=flash_attention_recompute,
+        flash_attention_fast_softmax=flash_attention_fast_softmax,
+        flash_attention_causal_mask=flash_attention_causal_mask,
     )
     attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
     outputs = attn_outputs[1:]
@@ -167,13 +330,21 @@ def gaudi_gpt_bigcode_model_forward(
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     token_idx: Optional[torch.Tensor] = None,
+    use_flash_attention: Optional[bool] = False,
+    flash_attention_recompute: Optional[bool] = False,
+    flash_attention_fast_softmax: Optional[bool] = False,
+    flash_attention_causal_mask: Optional[bool] = False,
 ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
     """
-    Copied from GPTBigCodeModel.forward: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
+    Copied from GPTBigCodeModel.forward: https://github.com/huggingface/transformers/blob/v4.40-release/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
     The only differences are:
-    - add new args token_idx
+    - add new args token_idx, use_flash_attention, flash_attention_recompute, flash_attention_fast_softmax, flash_attention_causal_mask
     - if token_idx and past_key_values are passed, set self_attention_mask based on the static shape of past_key_values
     """
+
+    # This flag used for correct tensors reshape for attention kernel
+    self._use_sdpa = use_flash_attention
+
     output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
     output_hidden_states = (
         output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
@@ -322,6 +493,10 @@ def gaudi_gpt_bigcode_model_forward(
                 use_cache=use_cache,
                 output_attentions=output_attentions,
                 token_idx=token_idx,
+                use_flash_attention=use_flash_attention,
+                flash_attention_recompute=flash_attention_recompute,
+                flash_attention_fast_softmax=flash_attention_fast_softmax,
+                flash_attention_causal_mask=flash_attention_causal_mask,
             )
 
         hidden_states = outputs[0]
@@ -358,10 +533,10 @@ def gaudi_gpt_bigcode_model_forward(
 
 class GaudiGPTBigCodeForCausalLM(GPTBigCodeForCausalLM):
     """
-    Inherits from GPTBigCodeForCausalLM: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
+    Inherits from GPTBigCodeForCausalLM: https://github.com/huggingface/transformers/blob/v4.40-release/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
     The only differences are:
-    - add new args token_idx
-    - add token_idx into model_inputs
+    - add new args token_idx, use_flash_attention, flash_attention_recompute, flash_attention_fast_softmax, flash_attention_causal_mask
+    - add token_idx, use_flash_attention, flash_attention_recompute, flash_attention_fast_softmax, flash_attention_causal_mask into model_inputs
     - when KV cache is enabled, slice next_input_ids from input_ids based on the token_idx
     - when KV cache is enabled, slice next_position_ids from position_ids based on the token_idx
     """
@@ -422,6 +597,10 @@ def prepare_inputs_for_generation(
                 "attention_mask": attention_mask,
                 "token_type_ids": token_type_ids,
                 "token_idx": token_idx,
+                "use_flash_attention": kwargs.get("use_flash_attention", False),
+                "flash_attention_recompute": kwargs.get("flash_attention_recompute", False),
+                "flash_attention_fast_softmax": kwargs.get("flash_attention_fast_softmax", False),
+                "flash_attention_causal_mask": kwargs.get("flash_attention_causal_mask", False),
             }
         )
         return model_inputs
@@ -443,6 +622,10 @@ def forward(
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         token_idx: Optional[torch.Tensor] = None,
+        use_flash_attention: Optional[bool] = False,
+        flash_attention_recompute: Optional[bool] = False,
+        flash_attention_fast_softmax: Optional[bool] = False,
+        flash_attention_causal_mask: Optional[bool] = False,
     ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
         r"""
         labels (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -467,6 +650,10 @@ def forward(
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
             token_idx=token_idx,
+            use_flash_attention=use_flash_attention,
+            flash_attention_recompute=flash_attention_recompute,
+            flash_attention_fast_softmax=flash_attention_fast_softmax,
+            flash_attention_causal_mask=flash_attention_causal_mask,
         )
         hidden_states = transformer_outputs[0]
 

diff --git a/tests/test_text_generation_example.py b/tests/test_text_generation_example.py
@@ -25,7 +25,7 @@
             ("EleutherAI/gpt-neox-20b", 1, False, 50.67672679310354),
             ("meta-llama/Llama-2-7b-hf", 1, True, 141.25776956002076),
             ("tiiuae/falcon-40b", 1, True, 25.202450111088346),
-            ("bigcode/starcoder", 1, False, 65.58632640700114),
+            ("bigcode/starcoder", 256, False, 4329.754794647058),
             ("Salesforce/codegen2-1B", 1, False, 446.4029486883532),
             ("mosaicml/mpt-30b", 1, False, 36.06464336116623),
             ("mistralai/Mistral-7B-v0.1", 1, True, 130.2172236767782),
@@ -142,6 +142,9 @@ def _test_text_generation(
     if "falcon" in model_name.lower() or "starcoder2" in model_name.lower():
         command += ["--use_flash_attention", "--flash_attention_causal_mask"]
 
+    if "starcoder" in model_name.lower() and "starcoder2" not in model_name.lower():
+        command += ["--use_flash_attention"]
+
     if "starcoder2" in model_name.lower():
         command += ["--flash_attention_recompute"]