Optimize for proper flash attn causal handling (#2503)

fastchat/train/llama2_flash_attn_monkey_patch.py

@@ -2,6 +2,7 @@
 from typing import Optional, Tuple
 
 import torch
+from flash_attn import __version__ as flash_attn_version
 from flash_attn.bert_padding import pad_input, unpad_input
 from flash_attn.flash_attn_interface import (
     flash_attn_func,
@@ -36,6 +37,7 @@ def forward(
     past_key_value: Optional[Tuple[torch.Tensor]] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
+    padding_mask: Optional[torch.Tensor] = None,
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
     if output_attentions:
         warnings.warn(
@@ -58,50 +60,31 @@ def forward(
     kv_seq_len = k.shape[1]
     past_kv_len = 0
     if past_key_value is not None:
-        past_kv_len = past_key_value[0].shape[1]
+        past_kv_len = past_key_value[0].shape[2]
         kv_seq_len += past_kv_len
 
     cos_sin = self.rotary_emb(v, seq_len=kv_seq_len)
     q, k = apply_rotary_pos_emb(q, k, cos_sin, position_ids)
 
     if past_key_value is not None:
+        assert (
+            flash_attn_version >= "2.1.0"
+        ), "past_key_value support requires flash-attn >= 2.1.0"
         # reuse k, v
-        k = torch.cat([past_key_value[0], k], dim=1)
-        v = torch.cat([past_key_value[1], v], dim=1)
-
-    past_key_value = (k, v) if use_cache else None
-
-    key_padding_mask = attention_mask
-    # Ideally we could just do this:
-    #  q, indices, cu_q_lens, max_s = unpad_input(q, key_padding_mask[:, -q_len:])
-    # but this does not work as Flash attention treats the q seq and kv seq as starting at index 0
-    # which then breaks the causality logic. Probably if q_len >> past_kv_len we should
-    # just skip flash attention. Leaving this in for now to demonstrate correctness of
-    # flash attention information even when q needs padding.
-    # TODO(siddartha): delegate back to original implementation on this condition.
-    if past_kv_len > 0:
-        q = torch.cat(
-            (
-                torch.full(
-                    (bsz, past_kv_len, self.num_heads, self.head_dim),
-                    0.0,
-                    dtype=q.dtype,
-                    device=q.device,
-                ),
-                q,
-            ),
-            dim=1,
-        )
+        k = torch.cat([past_key_value[0].transpose(1, 2), k], dim=1)
+        v = torch.cat([past_key_value[1].transpose(1, 2), v], dim=1)
+
+    past_key_value = (k.transpose(1, 2), v.transpose(1, 2)) if use_cache else None
 
-    if key_padding_mask is None:
+    if attention_mask is None:
         output = flash_attn_func(q, k, v, 0.0, softmax_scale=None, causal=True).view(
-            bsz, q_len + past_kv_len, -1
+            bsz, q_len, -1
         )
     else:
-        q, indices, cu_q_lens, max_s = unpad_input(q, key_padding_mask)
+        q, indices, cu_q_lens, max_s = unpad_input(q, attention_mask[:, -q_len:])
         # We can skip concat and call unpad twice but seems better to call unpad only once.
         kv, _, cu_k_lens, max_k = unpad_input(
-            torch.stack((k, v), dim=2), key_padding_mask
+            torch.stack((k, v), dim=2), attention_mask
         )
         output_unpad = flash_attn_varlen_kvpacked_func(
             q,
@@ -115,11 +98,7 @@ def forward(
             causal=True,
         )
         output_unpad = output_unpad.reshape(-1, self.num_heads * self.head_dim)
-        output = pad_input(output_unpad, indices, bsz, q_len + past_kv_len)
-
-    # Need to strip off the zero query outputs.
-    if past_kv_len > 0:
-        output = output[:, past_kv_len:, ...]
+        output = pad_input(output_unpad, indices, bsz, q_len)
 
     return self.o_proj(output), None, past_key_value
 
@@ -245,7 +224,7 @@ def test():
                 use_cache=True,
             )
             parts.append(part)
-            past_kv_len = past_kv[0].shape[1]
+            past_kv_len = past_kv[0].shape[2]
 
         print(
             f"allclose(oneshot[:, 0], parts[0]) = {torch.allclose(oneshot[:, :part_len], parts[0])}"