Enable FusedSDPA for prompt attention with VLLM_PROMPT_USE_FUSEDSDPA (#…

…168)

vllm/attention/backends/habana_attn.py

@@ -2,6 +2,7 @@
 # Copyright (C) 2024 Habana Labs, Ltd. an Intel Company
 ###############################################################################
 
+import os
 from dataclasses import dataclass
 from typing import Any, Dict, List, Optional, Tuple, Type
 
@@ -166,6 +167,12 @@ def __init__(
         assert self.num_heads % self.num_kv_heads == 0
         self.num_queries_per_kv = self.num_heads // self.num_kv_heads
 
+        self.prefill_usefusedsdpa = os.getenv('VLLM_PROMPT_USE_FUSEDSDPA',
+                                              '0').lower() in ['1', 'true']
+        if self.prefill_usefusedsdpa:
+            assert alibi_slopes is None, \
+                'Prefill with FusedSDPA not supported with alibi slopes!'
+
         suppored_head_sizes = HabanaPagedAttention.get_supported_head_sizes()
         if head_size not in suppored_head_sizes:
             raise ValueError(
@@ -223,15 +230,18 @@ def forward(
         if attn_metadata.is_prompt:
             # Prompt run.
             if kv_cache is None or attn_metadata.block_tables.numel() == 0:
-                # TODO: move this outside of model
-                assert attn_metadata.attn_bias is not None, \
-                       'attn_bias must be set before calling model.forward!'
-                attn_bias = attn_metadata.attn_bias
-                if self.alibi_slopes is not None and \
-                   self.position_bias is not None:
-                    attn_bias.add_(self.position_bias[:, :,
-                                                      -attn_bias.size(2):,
-                                                      -attn_bias.size(3):])
+                if not self.prefill_usefusedsdpa:
+                    # TODO: move this outside of model
+                    assert attn_metadata.attn_bias is not None, \
+                        'attn_bias must be set before calling model.forward!'
+                    attn_bias = attn_metadata.attn_bias
+                    if self.alibi_slopes is not None and \
+                        self.position_bias is not None:
+                        attn_bias.add_(self.position_bias[:, :,
+                                                          -attn_bias.size(2):,
+                                                          -attn_bias.size(3):])
+                else:
+                    attn_bias = None
 
                 query_shape = (batch_size, seq_len, self.num_heads,
                                self.head_size)
@@ -247,6 +257,7 @@ def forward(
                     matmul_qk_op=self.matmul_qk,
                     softmax_op=self.softmax,
                     matmul_av_op=self.matmul_av,
+                    valid_seq_lengths=attn_metadata.seq_lens_tensor,
                 )
                 output = out.reshape(batch_size, seq_len, hidden_size)
             else:

vllm/hpu/ops.py

@@ -21,6 +21,13 @@
 except ImportError:
     logger.warning("Could not import HPU FusedRMSNorm kernel. "
                    "vLLM will use forward_native implementation of RMSNorm.")
+HPUFusedSDPA = None
+try:
+    from habana_frameworks.torch.hpex.kernels import FusedSDPA
+    HPUFusedSDPA = FusedSDPA
+except ImportError:
+    logger.warning("Could not import HPU FusedSDPA kernel. "
+                   "vLLM will use native implementation.")
 
 PA_SPLIT_VALUE = (os.environ.get('PA_SPLIT_VALUE', '1') == '1')
 
@@ -126,6 +133,21 @@ def static_fused_moe(hidden_states, w1, w2, score, topk):
     return final_hidden_states.view(-1, D)
 
 
+#TODO: remove after fusedsdpa fix for query_head != kv_head
+def repeat_kv(kv: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep).
+    The kv go from (batch, num_key_value_heads, seqlen, head_dim) to
+    (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = kv.shape
+    if n_rep == 1:
+        return kv
+    kv = kv[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen,
+                                     head_dim)
+    return kv.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
 def prompt_attention(
     query: torch.Tensor,
     key: torch.Tensor,
@@ -136,24 +158,36 @@ def prompt_attention(
     matmul_qk_op=torch.matmul,
     softmax_op=torch.softmax,
     matmul_av_op=torch.matmul,
+    valid_seq_lengths: Optional[torch.Tensor] = None,
 ) -> torch.Tensor:
     query = query.transpose(1, 2)
     key = key.transpose(1, 2)
     value = value.transpose(1, 2)
     query_heads = query.size(1)
     kv_heads = key.size(1)
-    if query_heads != kv_heads:
-        query = query.unflatten(1, (kv_heads, -1))
-        key = key.unflatten(1, (kv_heads, 1))
-        value = value.unflatten(1, (kv_heads, 1))
+    if attn_bias is not None or HPUFusedSDPA is None:
+        if query_heads != kv_heads:
+            query = query.unflatten(1, (kv_heads, -1))
+            key = key.unflatten(1, (kv_heads, 1))
+            value = value.unflatten(1, (kv_heads, 1))
+            if attn_bias is not None:
+                attn_bias = attn_bias.unsqueeze(2)
+        attn_weights = matmul_qk_op(query * scale, key.transpose(-1, -2))
         if attn_bias is not None:
-            attn_bias = attn_bias.unsqueeze(2)
-    attn_weights = matmul_qk_op(query * scale, key.transpose(-1, -2))
-    if attn_bias is not None:
-        attn_weights.add_(attn_bias)
-    attn_weights = softmax_op(attn_weights, dim=-1)
-    attn_weights = matmul_av_op(attn_weights, value)
-    if query_heads != kv_heads:
-        attn_weights = attn_weights.flatten(1, 2)
+            attn_weights.add_(attn_bias)
+        attn_weights = softmax_op(attn_weights, dim=-1)
+        attn_weights = matmul_av_op(attn_weights, value)
+        if query_heads != kv_heads:
+            attn_weights = attn_weights.flatten(1, 2)
+    else:
+        #TODO: remove after fusedsdpa fix for query_heads != kv_heads
+        if query_heads != kv_heads:
+            key = repeat_kv(key, int(query_heads // kv_heads))
+            value = repeat_kv(value, int(query_heads // kv_heads))
+        softmax_mode = 'fast'
+        recompute_mode = True
+        attn_weights = FusedSDPA.apply(query, key, value, None, 0.0, True,
+                                       scale, softmax_mode, recompute_mode,
+                                       valid_seq_lengths, 'right')
     attn_weights = attn_weights.transpose(1, 2)
     return attn_weights

vllm/worker/habana_model_runner.py

@@ -151,6 +151,9 @@ class HpuModelAdapter():
 
     def __init__(self, model, enforce_eager):
         self.model = model
+        self.prefill_use_fusedsdpa = os.getenv('VLLM_PROMPT_USE_FUSEDSDPA',
+                                               '0').lower() in ['1', 'true']
+
         if not htorch.utils.internal.is_lazy() and not enforce_eager:
             self.model = torch.compile(self.model,
                                        backend='hpu_backend',
@@ -159,7 +162,7 @@ def __init__(self, model, enforce_eager):
     def _set_attn_bias(self, attn_metadata, batch_size, seq_len, device,
                        dtype):
         prefill_metadata = attn_metadata
-        if prefill_metadata is None:
+        if prefill_metadata is None or self.prefill_use_fusedsdpa:
             return attn_metadata
 
         seq_lens_t = prefill_metadata.seq_lens_tensor
@@ -599,7 +602,6 @@ def _prepare_prompt(
             # actual prompt lens
             context_lens.append(context_len)
             query_lens.append(seq_len - context_len)
-
             input_tokens.append(prompt_tokens)
             # NOTE(woosuk): Here we assume that the first token in the prompt
             # is always the first token in the sequence.
@@ -672,7 +674,6 @@ def _prepare_prompt(
         max_prompt_len = max(
             find_bucket(max(seq_lens), self.prompt_seq_bucket_cfg),
             self.block_size)
-
         input_tokens = make_tensor_with_pad(input_tokens,
                                             max_len=max_prompt_len,
                                             pad=0,