Drop the GLM-5.3-Flash breakable prefill chunk-size pin (#41837)
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
This commit is contained in:
co-authored by
Mohammad Angkad
parent
bb62f125aa
commit
99c59a2ea6
@@ -409,24 +409,6 @@ def disable_prefill_cuda_graph_for_deepseek_trtllm_mla(server_args: Any):
|
||||
)
|
||||
|
||||
|
||||
def apply_glm5_chunked_prefill_default(server_args: Any):
|
||||
"""Set the opted-in GLM BCG chunk default before memory budgeting."""
|
||||
cfg = resolving_view(server_args)
|
||||
if (
|
||||
get_platform().is_cuda
|
||||
and (Phase.PREFILL, "backend") in server_args._cuda_graph_config_locked
|
||||
and cfg.cuda_graph_config.prefill.backend == Backend.BREAKABLE
|
||||
and cfg.chunked_prefill_size is None
|
||||
and "Glm5NextForConditionalGeneration"
|
||||
in model_config_of(server_args).hf_config.architectures
|
||||
):
|
||||
declare_resolution(
|
||||
server_args,
|
||||
"_apply_glm5_chunked_prefill_default",
|
||||
chunked_prefill_size=4096,
|
||||
)
|
||||
|
||||
|
||||
def apply_glm5_prefill_cuda_graph_policy(server_args: Any):
|
||||
"""Set capture sizes for explicitly enabled GLM breakable prefill graphs."""
|
||||
cfg = resolving_view(server_args)
|
||||
|
||||
@@ -162,7 +162,6 @@ def run_resolution_pipeline(server_args: Any) -> None:
|
||||
# resolution (the declarative registry materializes too late to affect
|
||||
# it). Inkling opts into full-graph prefill capture here.
|
||||
from sglang.srt.arg_groups.cuda_graph_hook import (
|
||||
apply_glm5_chunked_prefill_default,
|
||||
apply_glm5_prefill_cuda_graph_policy,
|
||||
apply_inkling_prefill_cuda_graph_default,
|
||||
apply_muse_glimmer_prefill_cuda_graph_max_bs_default,
|
||||
@@ -178,9 +177,6 @@ def run_resolution_pipeline(server_args: Any) -> None:
|
||||
run_hook(handle_dwdp, server_args)
|
||||
|
||||
run_hook(handle_cuda_graph_config, server_args)
|
||||
# Requires the parsed backend and explicit-input locks, and must precede
|
||||
# handle_gpu_memory_settings so the chunk size feeds memory budgeting.
|
||||
run_hook(apply_glm5_chunked_prefill_default, server_args)
|
||||
|
||||
from sglang.srt.arg_groups.platform_hook import (
|
||||
handle_amd_specifics,
|
||||
|
||||
@@ -40,7 +40,6 @@ _OVERRIDABLE_HOOKS: FrozenSet[str] = frozenset(
|
||||
"apply_muse_glimmer_prefill_cuda_graph_max_bs_default",
|
||||
"handle_dwdp",
|
||||
"handle_cuda_graph_config",
|
||||
"apply_glm5_chunked_prefill_default",
|
||||
"handle_hpu_backends",
|
||||
"handle_cpu_backends",
|
||||
"handle_npu_backends",
|
||||
|
||||
Reference in New Issue
Block a user