[PyTorch] Pad V when Q/V head dims differ (MLA) for THD

Signed-off-by: Hollow Man <hollowman@opensuse.org>

Author: HollowMan6

transformer_engine/pytorch/attention/dot_product_attention/dot_product_attention.py

@@ -11,6 +11,7 @@
 import logging
 
 import torch
+import torch.nn.functional as F
 from torch.nn.parameter import Parameter
 
 import transformer_engine_torch as tex
@@ -179,6 +180,18 @@
 __all__ = ["DotProductAttention"]
 
 
+def _pad_thd_value_layer(value_layer, head_dim_qk):
+    """Pad V for THD FlashAttention when Q/K and V head dimensions differ."""
+    orig_head_dim_v = value_layer.shape[-1]
+    return F.pad(value_layer, (0, head_dim_qk - orig_head_dim_v)), orig_head_dim_v
+
+
+def _trim_thd_output(attn_out, num_attention_heads, padded_head_dim_v, orig_head_dim_v):
+    """Trim FlashAttention THD output after padding V to the Q/K head dimension."""
+    attn_out = attn_out.reshape(attn_out.shape[0], num_attention_heads, padded_head_dim_v)
+    return attn_out[..., :orig_head_dim_v].reshape(attn_out.shape[0], -1)
+
+
 class DotProductAttention(TransformerEngineBaseModule):
     r"""Allows the model to jointly attend to information from different
     representation subspaces as described in the paper:
@@ -1508,6 +1521,16 @@ def forward(
             )
 
             if use_flash_attention:
+                orig_v_dim = None
+                if (
+                    q_format == "thd"
+                    and kv_format == "thd"
+                    and not isinstance(value_layer, Float8TensorStorage)
+                    and head_dim_qk != head_dim_v
+                    and value_layer.shape[-1] < head_dim_qk
+                ):
+                    value_layer, orig_v_dim = _pad_thd_value_layer(value_layer, head_dim_qk)
+
                 if core_attention_bias_type == "alibi":
                     alibi_slopes, _ = dpa_utils.get_alibi(
                         _alibi_cache,
@@ -1516,7 +1539,7 @@ def forward(
                         max_seqlen_kv,
                         alibi_slopes=alibi_slopes,
                     )
-                return self.flash_attention(
+                attn_out = self.flash_attention(
                     query_layer,
                     key_layer,
                     value_layer,
@@ -1541,6 +1564,9 @@ def forward(
                     fp8_output=fp8_output,
                     num_splits=num_splits,
                 )
+                if orig_v_dim is not None:
+                    return _trim_thd_output(attn_out, num_attention_heads, head_dim_qk, orig_v_dim)
+                return attn_out
 
             if use_fused_attention:
                 fu_core_attention_bias_type = core_attention_bias_type

transformer_engine/pytorch/attention/dot_product_attention/utils.py

@@ -709,10 +709,6 @@ def get_attention_backend(
 
     # Filter: Head dimension
     if head_dim_qk != head_dim_v:
-        if use_flash_attention_2 and FlashAttentionUtils.is_installed:
-            logger.debug("Disabling FlashAttention 2 as it does not support MLA.")
-            use_flash_attention_2 = False
-
         qkv_layout_group = qkv_layout.replace("b", "").replace("s", "").replace("t", "")
         if use_fused_attention and qkv_layout_group != "hd_hd_hd":
             logger.debug(