Merge remote-tracking branch 'upstream/main' into patches
Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
@@ -410,17 +410,24 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:
|
||||
Without autotuning, FlashInfer will rely on heuristics, which may
|
||||
be significantly slower.
|
||||
|
||||
Every rank profiles the same tactics. When distributed, per-tactic
|
||||
timings are averaged over the world CPU group so all ranks select the
|
||||
same tactic.
|
||||
With PP > 1, stages run different layers and may profile different ops,
|
||||
so each stage's TP group tunes separately with its own cache file;
|
||||
otherwise the world group tunes together. Per-tactic timings are
|
||||
averaged over the tuning group so all its ranks select the same tactic.
|
||||
"""
|
||||
from flashinfer.autotuner import AutoTuner, set_autotune_process_group
|
||||
|
||||
import vllm.utils.flashinfer as fi_utils
|
||||
from vllm.distributed.parallel_state import get_world_group
|
||||
from vllm.distributed.parallel_state import (
|
||||
get_pp_group,
|
||||
get_tp_group,
|
||||
get_world_group,
|
||||
)
|
||||
|
||||
world = get_world_group()
|
||||
is_leader = world.rank_in_group == 0
|
||||
pp_size = get_pp_group().world_size
|
||||
tune_group = get_tp_group() if pp_size > 1 else world
|
||||
is_leader = tune_group.rank_in_group == 0
|
||||
tuner = AutoTuner.get()
|
||||
|
||||
autotune_kwargs: dict = {}
|
||||
@@ -433,6 +440,11 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:
|
||||
autotune_kwargs["skip_ops"] = skip_ops
|
||||
|
||||
cache_path = resolve_flashinfer_autotune_file(runner)
|
||||
if pp_size > 1:
|
||||
ranks = "-".join(str(rank) for rank in tune_group.ranks)
|
||||
cache_path = cache_path.with_name(
|
||||
f"{cache_path.stem}_tp_{ranks}{cache_path.suffix}"
|
||||
)
|
||||
if is_leader:
|
||||
logger.info_once("Using FlashInfer autotune cache file: %s", cache_path)
|
||||
|
||||
@@ -441,18 +453,18 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:
|
||||
# which lead to some EP ranks receiving no tokens and skipping their
|
||||
# MoE kernel entirely, and cause hang due to all-reduce collective
|
||||
# during synchronized autotuning.
|
||||
# Read cached autotune results and broadcast to all ranks.
|
||||
# Read cached autotune results and broadcast within the tuning group.
|
||||
cached_results: bytes | None = None
|
||||
if is_leader and cache_path.exists():
|
||||
with open(cache_path, "rb") as f:
|
||||
cached_results = f.read()
|
||||
cached_results = world.broadcast_object(cached_results, src=0)
|
||||
cached_results = tune_group.broadcast_object(cached_results, src=0)
|
||||
if cached_results is not None:
|
||||
write_flashinfer_autotune_cache(cache_path, cached_results)
|
||||
world.barrier()
|
||||
tune_group.barrier()
|
||||
tuner.load_configs(str(cache_path))
|
||||
|
||||
group = world.cpu_group if world.world_size > 1 else None
|
||||
group = tune_group.cpu_group if tune_group.world_size > 1 else None
|
||||
set_autotune_process_group(group)
|
||||
try:
|
||||
with (
|
||||
|
||||
@@ -4,6 +4,7 @@
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import sys
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
import torch
|
||||
@@ -19,7 +20,10 @@ logger = init_logger(__name__)
|
||||
|
||||
|
||||
def _get_kda_layer(worker: Worker) -> KimiK3DeltaAttention | None:
|
||||
from vllm.models.kimi_k3.nvidia.kda import KimiK3DeltaAttention
|
||||
# Kimi model construction already imports kda. Avoid importing it here.
|
||||
kda = sys.modules.get("vllm.models.kimi_k3.nvidia.kda")
|
||||
if kda is None:
|
||||
return None
|
||||
|
||||
compilation_config = getattr(
|
||||
worker.model_runner,
|
||||
@@ -33,7 +37,7 @@ def _get_kda_layer(worker: Worker) -> KimiK3DeltaAttention | None:
|
||||
(
|
||||
layer
|
||||
for layer in static_context.values()
|
||||
if isinstance(layer, KimiK3DeltaAttention)
|
||||
if isinstance(layer, kda.KimiK3DeltaAttention)
|
||||
),
|
||||
None,
|
||||
)
|
||||
|
||||
@@ -21,6 +21,7 @@ logger = init_logger(__name__)
|
||||
|
||||
def _replayssm_autotune_kwargs(
|
||||
runner: "GPUModelRunner",
|
||||
max_token_prefill_kwargs: dict[str, Any],
|
||||
) -> tuple[int, dict[str, Any]] | None:
|
||||
config = runner.vllm_config
|
||||
if not (
|
||||
@@ -46,11 +47,16 @@ def _replayssm_autotune_kwargs(
|
||||
runner.max_num_tokens // query_len,
|
||||
runner.kv_cache_config.num_blocks - 1,
|
||||
)
|
||||
if max_num_reqs <= 0:
|
||||
logger.warning_once(
|
||||
"Skipping FlashInfer ReplaySSM autotuning because no non-padding "
|
||||
"state slot is available."
|
||||
)
|
||||
return None
|
||||
|
||||
decode_kwargs = {
|
||||
**max_token_prefill_kwargs,
|
||||
"num_tokens": max_num_reqs * query_len,
|
||||
"skip_eplb": True,
|
||||
"is_profile": True,
|
||||
"randomize_inputs": True,
|
||||
"uniform_decode": True,
|
||||
}
|
||||
if config.use_v2_model_runner:
|
||||
@@ -75,18 +81,22 @@ def _temporary_replayssm_autotune_state(
|
||||
)
|
||||
|
||||
reset_tensors: dict[int, torch.Tensor] = {}
|
||||
tracker_specs: dict[int, tuple[torch.Tensor, torch.Tensor, int, int]] = {}
|
||||
tracker_specs: dict[
|
||||
int, tuple[torch.Tensor, torch.Tensor, torch.Tensor, int, int]
|
||||
] = {}
|
||||
for module in runner.get_model().modules():
|
||||
if not isinstance(module, MambaMixer2) or not module.use_replayssm:
|
||||
continue
|
||||
assert module.replayssm_buffer_len is not None
|
||||
ring_start = module._replayssm_ring_start
|
||||
prev_num_accepted = module._replayssm_prev_num_accepted
|
||||
prev_query_len = module._replayssm_prev_query_len
|
||||
tracker_specs.setdefault(
|
||||
ring_start.data_ptr(),
|
||||
(
|
||||
ring_start,
|
||||
prev_num_accepted,
|
||||
prev_query_len,
|
||||
module.replayssm_buffer_len,
|
||||
module.kv_cache[2].size(2),
|
||||
),
|
||||
@@ -95,6 +105,7 @@ def _temporary_replayssm_autotune_state(
|
||||
*module.kv_cache,
|
||||
ring_start,
|
||||
prev_num_accepted,
|
||||
prev_query_len,
|
||||
)
|
||||
for tensor in tensors:
|
||||
if tensor.numel():
|
||||
@@ -121,20 +132,26 @@ def _temporary_replayssm_autotune_state(
|
||||
for (
|
||||
ring_start,
|
||||
prev_num_accepted,
|
||||
prev_query_len,
|
||||
logical_window,
|
||||
ring_buffer_len,
|
||||
) in tracker_specs.values():
|
||||
# Compile reset (prefill) and advance (decode) before inference.
|
||||
# The final reset leaves the decode tuning run in a clean state.
|
||||
reset_replayssm_ring_trackers(ring_start, prev_num_accepted, state_slots)
|
||||
reset_replayssm_ring_trackers(
|
||||
ring_start, prev_num_accepted, prev_query_len, state_slots
|
||||
)
|
||||
update_replayssm_ring_trackers(
|
||||
ring_start,
|
||||
prev_num_accepted,
|
||||
prev_query_len,
|
||||
state_slots,
|
||||
logical_window,
|
||||
ring_buffer_len,
|
||||
)
|
||||
reset_replayssm_ring_trackers(ring_start, prev_num_accepted, state_slots)
|
||||
reset_replayssm_ring_trackers(
|
||||
ring_start, prev_num_accepted, prev_query_len, state_slots
|
||||
)
|
||||
|
||||
try:
|
||||
yield
|
||||
@@ -151,7 +168,13 @@ def _temporary_replayssm_autotune_state(
|
||||
|
||||
|
||||
def replayssm_autotune_warmup(runner: "GPUModelRunner") -> None:
|
||||
autotune = _replayssm_autotune_kwargs(runner)
|
||||
max_token_prefill_kwargs = {
|
||||
"num_tokens": runner.scheduler_config.max_num_batched_tokens,
|
||||
"skip_eplb": True,
|
||||
"is_profile": True,
|
||||
"randomize_inputs": True,
|
||||
}
|
||||
autotune = _replayssm_autotune_kwargs(runner, max_token_prefill_kwargs)
|
||||
if autotune is None:
|
||||
return
|
||||
max_num_reqs, decode_kwargs = autotune
|
||||
|
||||
Reference in New Issue
Block a user