option to check seq len at runtime

src/anemoi/models/layers/attention.py

@@ -72,25 +72,43 @@ def __init__(
         self.attention = attn_func
 
         self.resolution = resolution
+
+        self.use_document_masking = False
+        self.is_attn_compiled = False
+        self.compile_at_runtime=False
+        self.use_flex_Attn=False
 
         if _FLEX_ATTENTION_AVAILABLE and (os.environ.get("FLEX_ATTN", "") != "" ):
-            LOGGER.info("Using Flex attn")
-            #LOGGER.info(f"self.num_heads {self.num_heads} self.embed_dim {self.embed_dim} self.head_dim {self.head_dim} self.dropout {self.dropout_p}")
+            self.use_flex_Attn = True
 
-            if window_size != None:
-                def sliding_window(b, h, q_idx, kv_idx):
-                    return abs(q_idx - kv_idx) <= window_size
-
-                seq_len=calculate_seq_len(resolution=self.resolution)
-                LOGGER.debug(f"grid points = {seq_len} for {self.resolution} resolution")
-
-                # B and H can be None here because they are uniform, so the block mask can just be broadcast to these dims
-                #TODO check if B != 1, does it have to be set?
-                self.block_mask = create_block_mask(sliding_window, B=None, H=None, Q_LEN=seq_len, KV_LEN=seq_len,_compile=True)
-                self.attention = functools.partial(flex_attention, block_mask=self.block_mask) #Cache the block mask (attn blog post)
-            else:
-                self.attention = flex_attention
-            self.attention = compile(self.attention) #Must be compiled, otherwise entire seq_len^2 aray is materilised in memory -> OOM
+        if self.use_flex_Attn:
+            if not self.compile_at_runtime:
+                LOGGER.info("Using Flex attn")
+                #LOGGER.info(f"self.num_heads {self.num_heads} self.embed_dim {self.embed_dim} self.head_dim {self.head_dim} self.dropout {self.dropout_p}")
+
+                if self.use_document_masking:
+
+                    def document_causal_mask(b, h, q_idx, kv_idx):
+                        causal_mask = q_idx >= kv_idx
+                        document_mask = document_id[q_idx] == document_id[kv_idx]
+                        return causal_mask & document_mask
+
+
+                elif window_size != None:
+                    def sliding_window(b, h, q_idx, kv_idx):
+                        return abs(q_idx - kv_idx) <= window_size
+
+                    seq_len=calculate_seq_len(resolution=self.resolution)
+                    LOGGER.debug(f"grid points = {seq_len} for {self.resolution} resolution")
+
+                    # B and H can be None here because they are uniform, so the block mask can just be broadcast to these dims
+                    #TODO check if B != 1, does it have to be set?
+                    self.block_mask = create_block_mask(sliding_window, B=None, H=None, Q_LEN=seq_len, KV_LEN=seq_len,_compile=True)
+                    self.attention = functools.partial(flex_attention, block_mask=self.block_mask) #Cache the block mask (attn blog post)
+                else:
+                    self.attention = flex_attention
+                self.attention = compile(self.attention) #Must be compiled, otherwise entire seq_len^2 aray is materilised in memory -> OOM
+                self.is_attn_compiled=True
 
             if (self.is_causal):
                 LOGGER.error("Causal not yet supported when using flex_attn (but this would be an easy add). Please rerun with 'is_causal = False'")
@@ -121,7 +139,18 @@ def forward(
             )
             for t in (query, key, value)
         )
-
+
+
+        if (self.use_flex_Attn and not self.is_attn_compiled):
+            LOGGER.info("Compiling Flex Attn at runtime")
+            seq_len=x.shape[0]
+            def sliding_window(b, h, q_idx, kv_idx):
+                return abs(q_idx - kv_idx) <= self.window_size[0]
+            self.block_mask = create_block_mask(sliding_window, B=None, H=None, Q_LEN=seq_len, KV_LEN=seq_len,_compile=True)
+            self.attention = functools.partial(flex_attention, block_mask=self.block_mask) #Cache the block mask (attn blog post)
+            self.attention = compile(self.attention)
+            self.is_attn_compiled = True
+
         query = shard_heads(query, shapes=shapes, mgroup=model_comm_group)
         key = shard_heads(key, shapes=shapes, mgroup=model_comm_group)
         value = shard_heads(value, shapes=shapes, mgroup=model_comm_group)