[Model] add GLM-5.3-Flash support (#53906)

This commit is contained in:
Jiangyun Zhu
2026-09-03 09:40:35 -07:00
committed by GitHub
parent da8ec28268
commit 98ed0856f3
94 changed files with 17927 additions and 266 deletions
+97 -11
View File
@@ -1,5 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import inspect
import torch
# this import will also register the custom ops
@@ -29,6 +31,67 @@ HAS_TILELANG_MHC = _has_tilelang_mhc()
HAS_AITER_MHC = is_aiter_found_and_supported()
def _has_aiter_mhc_fused() -> bool:
if not HAS_AITER_MHC:
return False
try:
from aiter.ops.mhc import mhc_fused_post_pre
except Exception:
return False
return callable(mhc_fused_post_pre)
def _aiter_mhc_op_accepts_norm(op_name: str) -> bool:
if not HAS_AITER_MHC:
return False
try:
from aiter.ops import mhc as aiter_mhc
for candidate_name in (op_name, f"{op_name}_fake"):
op = getattr(aiter_mhc, candidate_name, None)
if op is None:
continue
parameters = inspect.signature(op).parameters
if "norm_weight" in parameters and "norm_eps" in parameters:
return True
except (AttributeError, ImportError, TypeError, ValueError):
pass
return False
HAS_AITER_MHC_FUSED = _has_aiter_mhc_fused()
HAS_AITER_MHC_PRE_NORM = _aiter_mhc_op_accepts_norm("mhc_pre")
HAS_AITER_MHC_FUSED_NORM = _aiter_mhc_op_accepts_norm("mhc_fused_post_pre")
def _aiter_mhc_supported(
residual: torch.Tensor,
norm_weight: torch.Tensor | None,
*,
supports_norm: bool,
) -> bool:
hidden_size = residual.shape[-1]
hc_mult = residual.shape[-2]
return (
HAS_AITER_MHC
and hidden_size % 256 == 0
and hc_mult == 4
and (norm_weight is None or supports_norm)
)
def _apply_mhc_norm(
layer_input: torch.Tensor,
norm_weight: torch.Tensor | None,
norm_eps: float,
) -> torch.Tensor:
if norm_weight is None:
return layer_input
from vllm import ir
return ir.ops.rms_norm(layer_input, norm_weight, norm_eps)
# --8<-- [start:mhc_pre]
@CustomOp.register("mhc_pre")
class MHCPreOp(CustomOp):
@@ -89,9 +152,11 @@ class MHCPreOp(CustomOp):
norm_weight: torch.Tensor | None = None,
norm_eps: float = 0.0,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
hidden_size = residual.shape[-1]
hc_mult = residual.shape[-2]
if HAS_AITER_MHC and hidden_size % 256 == 0 and hc_mult == 4:
if _aiter_mhc_supported(
residual,
norm_weight,
supports_norm=HAS_AITER_MHC_PRE_NORM,
):
return torch.ops.vllm.mhc_pre_aiter(
residual,
fn,
@@ -122,7 +187,7 @@ class MHCPreOp(CustomOp):
norm_eps,
)
else:
return self.forward_native(
post_mix, comb_mix, layer_input = self.forward_native(
residual,
fn,
hc_scale,
@@ -136,6 +201,11 @@ class MHCPreOp(CustomOp):
norm_weight,
norm_eps,
)
return (
post_mix,
comb_mix,
_apply_mhc_norm(layer_input, norm_weight, norm_eps),
)
def forward_native(
self,
@@ -225,9 +295,7 @@ class MHCPostOp(CustomOp):
post_layer_mix: torch.Tensor,
comb_res_mix: torch.Tensor,
) -> torch.Tensor:
hidden_size = residual.shape[-1]
hc_mult = residual.shape[-2]
if HAS_AITER_MHC and hidden_size % 256 == 0 and hc_mult == 4:
if _aiter_mhc_supported(residual, None, supports_norm=True):
return torch.ops.vllm.mhc_post_aiter(
x,
residual,
@@ -449,9 +517,11 @@ class MHCFusedPostPreOp(CustomOp):
norm_weight: torch.Tensor | None = None,
norm_eps: float = 0.0,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
hidden_size = residual.shape[-1]
hc_mult = residual.shape[-2]
if HAS_AITER_MHC and hidden_size % 256 == 0 and hc_mult == 4:
if HAS_AITER_MHC_FUSED and _aiter_mhc_supported(
residual,
norm_weight,
supports_norm=HAS_AITER_MHC_FUSED_NORM,
):
return torch.ops.vllm.mhc_fused_post_pre_aiter(
x,
residual,
@@ -489,7 +559,7 @@ class MHCFusedPostPreOp(CustomOp):
norm_weight,
norm_eps,
)
return self.forward_native(
residual_cur, post_mix_cur, comb_mix_cur, layer_input_cur = self.forward_native(
x,
residual,
post_layer_mix,
@@ -507,6 +577,12 @@ class MHCFusedPostPreOp(CustomOp):
norm_weight,
norm_eps,
)
return (
residual_cur,
post_mix_cur,
comb_mix_cur,
_apply_mhc_norm(layer_input_cur, norm_weight, norm_eps),
)
def forward_native(
self,
@@ -577,3 +653,13 @@ class MHCFusedPostPreOp(CustomOp):
hc_post_mult_value,
sinkhorn_repeat,
)
def hc_expand(x: torch.Tensor, n: int) -> torch.Tensor:
"""[s, hidden_size] -> [s, n * hidden_size] by replication."""
return x.unsqueeze(1).expand(-1, n, -1).contiguous()
def hc_contract(x: torch.Tensor, n: int) -> torch.Tensor:
"""[s, n * hidden_size] -> [s, hidden_size] by averaging."""
return x.mean(dim=1)