minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Revert "[Spec] Split init_backends; account draft weights in --mem-fraction-static" (#28841)

Browse Source
This commit is contained in:
Lianmin Zheng
2026-06-21 07:52:35 -07:00
committed by GitHub
parent 320b231ea6
commit 7942d546d1
12 changed files with 99 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/model_runner.py
+33 -23
View File
@@ -788,6 +788,18 @@ class ModelRunner(ModelRunnerKVCacheMixin):
# Deduce KV cache dtype
self.configure_kv_cache_dtype()
# Snapshot free memory at the end of the weight-load phase. KV-pool
# profiling uses this instead of measuring at alloc_memory_pool()
# time: draft-model weights load between the two phases and must stay
# outside the --mem-fraction-static budget (deployments tune the
# fraction assuming draft weights live in the non-static slack).
self.post_model_load_memory = get_available_gpu_memory(
self.device,
self.gpu_id,
distributed=get_world_group().world_size > 1,
cpu_group=get_world_group().cpu_group,
)
def get_pp_proxy_topk_size(self) -> Optional[int]:
hf_config = self.model_config.hf_text_config
if (
@@ -850,13 +862,18 @@ class ModelRunner(ModelRunnerKVCacheMixin):
self.graph_mem_usage = 0
self.prefill_cuda_graph_runner = None
def init_attention_backends(self):
"""Initialize attention backends only (no cuda graph capture)."""
def init_backends(self, disable_cuda_graph: bool = False):
"""Initialize attention backends and capture cuda graphs."""
server_args = self.server_args
# TODO: Refactor device-specific init branches into platform interface (separate PR).
# Must be called BEFORE init_decode_cuda_graph() so CUDA graph capture
# runs with aux hidden state capture enabled.
self.init_aux_hidden_state_capture()
# Device-specific attention-backend init. cg_supported gates decode
# cuda-graph capture (out-of-tree / unknown platforms may not support it).
cg_supported = True
if self.device == "cuda" or self.device == "musa":
self.init_cublas()
self.init_attention_backend()
@@ -875,15 +892,12 @@ class ModelRunner(ModelRunnerKVCacheMixin):
get_world_group().world_size,
get_world_group().cpu_group,
)
elif current_platform.is_out_of_tree():
self.init_attention_backend()
cg_supported = current_platform.support_cuda_graph()
else:
self.init_attention_backend()
def init_cuda_graphs(self, capture_decode_cuda_graph: bool = True):
"""Capture cuda graphs. Requires init_attention_backends() to have run.
Spec draft runners pass capture_decode_cuda_graph=False
because they capture their own decode-style graphs separately.
"""
cg_supported = False
# The eager (no-cuda-graph) phase runner, built AFTER the attention
# backend so its __init__ can warm up kernels (run-once) and allocate the
@@ -898,27 +912,23 @@ class ModelRunner(ModelRunnerKVCacheMixin):
# eager buffer allocated above. (init_prefill_cuda_graph routes prefill
# to the eager runner when the prefill graph is disabled.)
self.init_prefill_cuda_graph()
self.decode_cuda_graph_runner = None
self.graph_mem_usage = 0
if capture_decode_cuda_graph:
if self.device in ("cuda", "musa", "cpu", "npu"):
self.init_decode_cuda_graph()
elif (
current_platform.is_out_of_tree()
and current_platform.support_cuda_graph()
):
self.init_decode_cuda_graph()
if not disable_cuda_graph and cg_supported:
self.init_decode_cuda_graph()
else:
self.decode_cuda_graph_runner = None
self.graph_mem_usage = 0
if disable_cuda_graph:
# Decode cuda graph disabled: route eager decode through the
# EagerRunner (the dispatch gate isinstance(..., EagerRunner) keeps
# _forward_raw off any replay branch).
self.decode_cuda_graph_runner = self.eager_runner
# Register forward hooks AFTER cuda-graph capture so their tensor ops are
# not traced into any captured graph — capture stays hook-free and hooks
# fire only on the eager forward path (capture replay never runs Python
# hooks anyway).
if self.server_args.forward_hooks:
register_forward_hooks(self.model, self.server_args.forward_hooks)
if server_args.forward_hooks:
register_forward_hooks(self.model, server_args.forward_hooks)
self.prealloc_symmetric_memory_pool()
python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py
+18 -23
View File
@@ -82,29 +82,24 @@ _is_hip = is_hip()
class ModelRunnerKVCacheMixin:
def _profile_available_bytes(self: ModelRunner, pre_model_load_memory: int) -> int:
# KV pool budget = currently-free GPU memory minus the non-static runtime
# slack (pre_model_load_memory * (1 - mem_fraction_static)). Whatever is
# already resident (model weights, etc.) is thus charged against it.
available_gpu_memory = get_available_gpu_memory(
self.device,
self.gpu_id,
distributed=get_world_group().world_size > 1,
cpu_group=get_world_group().cpu_group,
)
# Use the snapshot taken at the end of this runner's weight-load phase,
# not the current free memory: draft-model weights loaded after that
# point are charged to the non-static slack, not the static budget.
post_model_load_memory = getattr(self, "post_model_load_memory", None)
if post_model_load_memory is None:
post_model_load_memory = get_available_gpu_memory(
self.device,
self.gpu_id,
distributed=get_world_group().world_size > 1,
cpu_group=get_world_group().cpu_group,
)
rest_memory = available_gpu_memory - pre_model_load_memory * (
rest_memory = post_model_load_memory - pre_model_load_memory * (
1 - self.mem_fraction_static
)
if self.mambaish_config is not None:
rest_memory = self.handle_max_mamba_cache(rest_memory)
# Loaded weights (target + draft) can exceed the static budget
if rest_memory <= 0:
raise ValueError(
f"Loaded weights leave no GPU memory for the KV cache under "
f"--mem-fraction-static={self.mem_fraction_static}."
)
return int(rest_memory * (1 << 30)) # return in bytes
def handle_max_mamba_cache(self: ModelRunner, total_rest_memory):
@@ -565,9 +560,9 @@ class ModelRunnerKVCacheMixin:
self.model_config.hf_text_config.swa_num_key_value_heads
// get_attention_tp_size(),
),
"swa_head_dim": self.model_config.swa_head_dim,
"swa_v_head_dim": self.model_config.swa_v_head_dim,
"v_head_dim": self.model_config.v_head_dim,
"swa_head_dim": self.model_config.hf_text_config.swa_head_dim,
"swa_v_head_dim": self.model_config.hf_text_config.swa_v_head_dim,
"v_head_dim": self.model_config.hf_text_config.v_head_dim,
}
self.token_to_kv_pool = SWAKVPool(
size=self.full_max_total_num_tokens,
@@ -688,9 +683,9 @@ class ModelRunnerKVCacheMixin:
self.model_config.hf_text_config.swa_num_key_value_heads
// get_attention_tp_size(),
),
"swa_head_dim": self.model_config.swa_head_dim,
"swa_v_head_dim": self.model_config.swa_v_head_dim,
"v_head_dim": self.model_config.v_head_dim,
"swa_head_dim": self.model_config.hf_text_config.swa_head_dim,
"swa_v_head_dim": self.model_config.hf_text_config.swa_v_head_dim,
"v_head_dim": self.model_config.hf_text_config.v_head_dim,
}
self.token_to_kv_pool = SWAKVPool(
size=self.full_max_total_num_tokens,