config: the DP/EP topology reads come from the parallel bag (#35025)

This commit is contained in:
Cheng Wan
2026-08-17 16:17:26 -07:00
committed by GitHub
parent d2bc697396
commit a97bc8db32
33 changed files with 332 additions and 304 deletions
@@ -240,10 +240,10 @@ def build_kv_cache(
retraction_backup = resolve_decode_retraction_backup(tp_worker=tp_worker)
disable_radix_cache = server_args.disable_radix_cache or (
disable_radix_cache = get_memory().disable_radix_cache or (
model_config.is_multimodal and uses_transformers_backend
)
if disable_radix_cache and not server_args.disable_radix_cache:
if disable_radix_cache and not get_memory().disable_radix_cache:
logger.warning(
"Radix cache is disabled for multimodal models with the "
"Transformers backend to avoid multimodal prefix-cache mismatches."
@@ -253,8 +253,8 @@ def build_kv_cache(
# these use specialized memory pools incompatible with the
# prefix-match-and-lock allocation path.
if (
server_args.disaggregation_decode_enable_radix_cache
and server_args.disaggregation_mode == "decode"
get_disagg().disaggregation_decode_enable_radix_cache
and get_disagg().disaggregation_mode == "decode"
):
if is_hybrid_swa:
raise ValueError(
@@ -285,15 +285,15 @@ def build_kv_cache(
),
is_eagle=spec_algorithm.is_eagle(),
tp_cache_group=(
attn_tp_cpu_group if server_args.enable_dp_attention else tp_cpu_group
attn_tp_cpu_group if get_parallel().enable_dp_attention else tp_cpu_group
),
attn_cp_cache_group=attn_cp_cpu_group,
attn_tp_cache_group=attn_tp_cpu_group,
pp_cache_group=pp_group.cpu_group,
eviction_policy=server_args.radix_eviction_policy,
eviction_policy=get_memory().radix_eviction_policy,
enable_metrics=enable_metrics,
enable_kv_cache_events=enable_kv_cache_events,
enable_session_radix_cache=server_args.enable_session_radix_cache,
enable_session_radix_cache=get_memory().enable_session_radix_cache,
enable_mamba_extra_buffer=server_args.enable_mamba_extra_buffer(),
enable_mamba_extra_buffer_lazy=server_args.enable_mamba_extra_buffer_lazy(),
pp_rank=ps.pp_rank,