[Kernel] Remove unused fake implementation (#55535)
Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
@@ -546,20 +546,10 @@ def olmo_hybrid_gdn_full_forward(
|
||||
)
|
||||
|
||||
|
||||
def olmo_hybrid_gdn_full_forward_fake(
|
||||
hidden_states: torch.Tensor,
|
||||
output: torch.Tensor,
|
||||
layer_name: str,
|
||||
) -> None:
|
||||
"""Fake implementation for torch.compile."""
|
||||
return
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="olmo_hybrid_gdn_full_forward",
|
||||
op_func=olmo_hybrid_gdn_full_forward,
|
||||
mutates_args=["output"],
|
||||
fake_impl=olmo_hybrid_gdn_full_forward_fake,
|
||||
)
|
||||
|
||||
|
||||
|
||||
@@ -1949,23 +1949,10 @@ def qwen_gdn_attention_core(
|
||||
)
|
||||
|
||||
|
||||
def gdn_attention_core_fake(
|
||||
qkv_or_qkvz: torch.Tensor,
|
||||
b_or_ba: torch.Tensor,
|
||||
a_or_z_out: torch.Tensor,
|
||||
core_attn_out: torch.Tensor,
|
||||
layer_name: LayerNameType,
|
||||
use_aiter: bool = False,
|
||||
) -> None:
|
||||
"""Fake implementation for torch.compile."""
|
||||
return
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="qwen_gdn_attention_core",
|
||||
op_func=qwen_gdn_attention_core,
|
||||
mutates_args=["a_or_z_out", "core_attn_out"],
|
||||
fake_impl=gdn_attention_core_fake,
|
||||
)
|
||||
|
||||
|
||||
@@ -1985,20 +1972,10 @@ def qwen_gdn_attention_core_fused_norm_packed(
|
||||
)
|
||||
|
||||
|
||||
def gdn_attention_core_fused_norm_packed_fake(
|
||||
mixed_qkvz: torch.Tensor,
|
||||
ba: torch.Tensor,
|
||||
core_attn_out: torch.Tensor,
|
||||
layer_name: LayerNameType,
|
||||
) -> None:
|
||||
return
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="qwen_gdn_attention_core_fused_norm_packed",
|
||||
op_func=qwen_gdn_attention_core_fused_norm_packed,
|
||||
mutates_args=["core_attn_out"],
|
||||
fake_impl=gdn_attention_core_fused_norm_packed_fake,
|
||||
)
|
||||
|
||||
|
||||
|
||||
@@ -336,18 +336,8 @@ def linear_attention(
|
||||
self._forward(hidden_states=hidden_states, output=output, positions=positions)
|
||||
|
||||
|
||||
def linear_attention_fake(
|
||||
hidden_states: torch.Tensor,
|
||||
output: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
layer_name: str,
|
||||
) -> None:
|
||||
return
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="linear_attention",
|
||||
op_func=linear_attention,
|
||||
mutates_args=["output"],
|
||||
fake_impl=linear_attention_fake,
|
||||
)
|
||||
|
||||
@@ -532,17 +532,8 @@ def mamba_mixer(
|
||||
self.forward_impl(hidden_states=hidden_states, output=output)
|
||||
|
||||
|
||||
def mamba_mixer_fake(
|
||||
hidden_states: torch.Tensor,
|
||||
output: torch.Tensor,
|
||||
layer_name: LayerNameType,
|
||||
) -> None:
|
||||
return
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="mamba_mixer",
|
||||
op_func=mamba_mixer,
|
||||
mutates_args=["output"],
|
||||
fake_impl=mamba_mixer_fake,
|
||||
)
|
||||
|
||||
@@ -1271,17 +1271,8 @@ def mamba_mixer2(
|
||||
self.conv_ssm_forward(projected_states=projected_states, output=output)
|
||||
|
||||
|
||||
def mamba_mixer2_fake(
|
||||
projected_states: torch.Tensor,
|
||||
output: torch.Tensor,
|
||||
layer_name: LayerNameType,
|
||||
) -> None:
|
||||
return
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="mamba_mixer2",
|
||||
op_func=mamba_mixer2,
|
||||
mutates_args=["output"],
|
||||
fake_impl=mamba_mixer2_fake,
|
||||
)
|
||||
|
||||
@@ -718,17 +718,6 @@ def _spec_aware_nonspec_subset(
|
||||
return attn_out.squeeze(0)
|
||||
|
||||
|
||||
def cpu_gdn_attention_core_fake(
|
||||
mixed_qkv: torch.Tensor,
|
||||
b: torch.Tensor,
|
||||
a: torch.Tensor,
|
||||
core_attn_out: torch.Tensor,
|
||||
layer_name: LayerNameType,
|
||||
) -> None:
|
||||
"""Fake implementation for torch.compile."""
|
||||
return
|
||||
|
||||
|
||||
def register_cpu_gdn_attention_ops() -> None:
|
||||
global _CPU_GDN_ATTENTION_OPS_REGISTERED
|
||||
if _CPU_GDN_ATTENTION_OPS_REGISTERED:
|
||||
@@ -738,6 +727,5 @@ def register_cpu_gdn_attention_ops() -> None:
|
||||
op_name="cpu_gdn_attention_core",
|
||||
op_func=cpu_gdn_attention_core,
|
||||
mutates_args=["core_attn_out"],
|
||||
fake_impl=cpu_gdn_attention_core_fake,
|
||||
)
|
||||
_CPU_GDN_ATTENTION_OPS_REGISTERED = True
|
||||
|
||||
@@ -373,17 +373,8 @@ def short_conv(
|
||||
self.forward_native(hidden_states=hidden_states, output=output)
|
||||
|
||||
|
||||
def short_conv_fake(
|
||||
hidden_states: torch.Tensor,
|
||||
output: torch.Tensor,
|
||||
layer_name: str,
|
||||
) -> None:
|
||||
return
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="short_conv",
|
||||
op_func=short_conv,
|
||||
mutates_args=["output"],
|
||||
fake_impl=short_conv_fake,
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user