[feat] QAD 5090: env-gate attention torch.compile via FASTVIDEO_DISABLE_ATTENTION_COMPILE

DistributedAttention.forward (and the VSA subclass) are hard-decorated with
@torch.compiler.disable, which keeps attention out of the surrounding
torch.compile graph unconditionally. That blocks the inference compile path
even after the FP4 linear and SageAttention3 graph-break fixes land, since
the attention forward itself can never be traced.

Make the disable conditional on FASTVIDEO_DISABLE_ATTENTION_COMPILE:
- unset / "1" / "true" (default): keep torch.compiler.disable — current behavior
- "0" / "false" / "no" / "off": drop it so attention can fold into the graph

The env var is read at import time (decorators are applied at class
definition), which is the right granularity for the multiproc spawn path:
each worker re-imports and inherits the parent's env.

Co-authored-by: Loay Rashid <42599591+loaydatrain@users.noreply.github.com>
Co-authored-by: Kaiqin Kong <k1kong@ucsd.edu>
Co-authored-by: William Lin <SolitaryThinker@users.noreply.github.com>

Author: 4 people

fastvideo/attention/layer.py

@@ -1,5 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
+import os
+
 import torch
 import torch.nn as nn
 
@@ -13,6 +15,26 @@
 from fastvideo.layers.rotary_embedding import _apply_rotary_emb
 
 
+def _attention_compile_disabled() -> bool:
+    """Whether to keep attention ``forward`` out of the torch.compile graph.
+
+    Defaults to ``True`` (the historical behavior: attention runs eager via
+    ``torch.compiler.disable``). Set ``FASTVIDEO_DISABLE_ATTENTION_COMPILE=0``
+    to let attention be traced/compiled into the surrounding graph.
+    """
+    val = os.environ.get("FASTVIDEO_DISABLE_ATTENTION_COMPILE")
+    if val is None:
+        return True
+    return val.strip().lower() not in ("0", "false", "no", "off", "")
+
+
+def _maybe_compiler_disable(fn):
+    """Apply ``torch.compiler.disable`` unless disabled via env var."""
+    if _attention_compile_disabled():
+        return torch.compiler.disable(fn)
+    return fn
+
+
 class DistributedAttention(nn.Module):
     """Distributed attention layer.
     """
@@ -56,7 +78,7 @@ def __init__(self,
         self.backend = backend_name_to_enum(attn_backend.get_name())
         self.dtype = dtype
 
-    @torch.compiler.disable
+    @_maybe_compiler_disable
     def forward(
         self,
         q: torch.Tensor,
@@ -146,7 +168,7 @@ class DistributedAttention_VSA(DistributedAttention):
     """Distributed attention layer with VSA support.
     """
 
-    @torch.compiler.disable
+    @_maybe_compiler_disable
     def forward(
         self,
         q: torch.Tensor,