[diffusion] Enable lossless Cosmos3 Super T2I QK fusion on Hopper TP2 (#40486)
This commit is contained in:
@@ -88,12 +88,11 @@ def _can_enable_t1_fused_qk_norm_rope(
|
||||
return False
|
||||
if is_blackwell:
|
||||
return True
|
||||
return (
|
||||
is_hopper
|
||||
and (hidden_act != "relu2" or hidden_size == 2048)
|
||||
and tp_size == 1
|
||||
and sp_size == 1
|
||||
)
|
||||
if not is_hopper or sp_size != 1:
|
||||
return False
|
||||
if tp_size == 1:
|
||||
return hidden_act != "relu2" or hidden_size == 2048
|
||||
return tp_size == 2 and hidden_act == "silu" and hidden_size == 5120
|
||||
|
||||
|
||||
# -----------------------------------------------------------------------------
|
||||
@@ -1728,8 +1727,8 @@ class Cosmos3OmniTransformer(CachableDiT, LayerwiseOffloadableModuleMixin):
|
||||
|
||||
# The T=1 fused path is faster on Blackwell. It also benefits the
|
||||
# single-GPU Hopper Nano (SwiGLU) and Edge (2048-wide dense) workloads,
|
||||
# while the Hopper Cosmos3-Super (dense MLP) multi-GPU workload remains on the split
|
||||
# path because that shape regresses with the fusion.
|
||||
# as well as TP2 Super-Text2Image (SwiGLU). The older dense-MLP Super
|
||||
# workload remains on the split path because that shape regresses.
|
||||
enable_t1_fused_qk_norm_rope = T == 1 and _can_enable_t1_fused_qk_norm_rope(
|
||||
is_blackwell=current_platform.is_blackwell(),
|
||||
is_hopper=current_platform.is_hopper(),
|
||||
|
||||
@@ -140,6 +140,24 @@ class TestCosmos3T1FusedQKNormRoPE(unittest.TestCase):
|
||||
def test_hopper_dense_mlp_disabled(self):
|
||||
self.assertFalse(self._can_enable(is_hopper=True, hidden_act="relu2"))
|
||||
|
||||
def test_hopper_super_t2i_tp2_enabled(self):
|
||||
self.assertTrue(self._can_enable(is_hopper=True, hidden_size=5120, tp_size=2))
|
||||
|
||||
def test_hopper_super_t2i_unsupported_topologies_disabled(self):
|
||||
for overrides in (
|
||||
{"hidden_act": "relu2"},
|
||||
{"hidden_act": "gelu"},
|
||||
{"hidden_size": 4096},
|
||||
{"tp_size": 4},
|
||||
{"sp_size": 2},
|
||||
{"is_compiled": True},
|
||||
{"is_hopper": False},
|
||||
):
|
||||
with self.subTest(overrides=overrides):
|
||||
settings = dict(is_hopper=True, hidden_size=5120, tp_size=2)
|
||||
settings.update(overrides)
|
||||
self.assertFalse(self._can_enable(**settings))
|
||||
|
||||
def test_hopper_edge_single_gpu_enabled(self):
|
||||
self.assertTrue(
|
||||
self._can_enable(is_hopper=True, hidden_act="relu2", hidden_size=2048)
|
||||
@@ -162,7 +180,7 @@ class TestCosmos3T1FusedQKNormRoPE(unittest.TestCase):
|
||||
self._can_enable(is_hopper=True, hidden_act="relu2", hidden_size=4096)
|
||||
)
|
||||
|
||||
def test_hopper_tensor_parallel_disabled(self):
|
||||
def test_hopper_other_tensor_parallel_shapes_disabled(self):
|
||||
self.assertFalse(self._can_enable(is_hopper=True, tp_size=2))
|
||||
|
||||
def test_hopper_sequence_parallel_disabled(self):
|
||||
|
||||
Reference in New Issue
Block a user