[Performance][MiniCPM-o] Whole-Euler CUDA graphs for Code2Wav with a shared attention arena (#8007)
This commit is contained in:
1 parent
0173cb3748
commit
3e62bba3cb
15 files changed
+3425
-210
No files matched your search
@@ -114,6 +114,7 @@ def _patched_estimator_step(
|
||||
attn_mask=None,
|
||||
valid_lengths=None,
|
||||
valid_frames=None,
|
||||
time_embedding=None,
|
||||
):
|
||||
assert _original_estimator_step is not None
|
||||
graph_runner = _backend_graph_runners.get(self)
|
||||
@@ -137,6 +138,7 @@ def _patched_estimator_step(
|
||||
attn_mask=attn_mask,
|
||||
valid_lengths=valid_lengths,
|
||||
valid_frames=valid_frames,
|
||||
time_embedding=time_embedding,
|
||||
)
|
||||
if (cnn_cache is None) != (att_cache is None):
|
||||
raise ValueError("estimator CNN and attention caches must both be present or absent")
|
||||
@@ -148,7 +150,8 @@ def _patched_estimator_step(
|
||||
|
||||
# The upstream embedder creates a frequency tensor on the host. Keep it
|
||||
# outside capture while retaining the tensor-only estimator body in graph.
|
||||
time_embedding = estimator.t_embedder(time).unsqueeze(1)
|
||||
if time_embedding is None:
|
||||
time_embedding = estimator.t_embedder(time).unsqueeze(1)
|
||||
if cnn_cache is None:
|
||||
return graph_runner.run(
|
||||
"cfm_estimator",
|
||||
|
||||
Reference in new issue
Block a user