[Core] MRV2 support for fast-prefill (#56145)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-09-10 10:36:04 -07:00
committed by GitHub
parent 7470082f57
commit e6cb56337b
12 changed files with 233 additions and 12 deletions
@@ -52,12 +52,23 @@ use_fork_for_test = (
@use_fork_for_test
@pytest.mark.parametrize("kv_sharing_fast_prefill", [False, True])
@pytest.mark.parametrize(
"kv_sharing_fast_prefill,use_v2_model_runner",
[
(False, False),
(True, False),
(True, True),
# (False, True) omitted: fast-prefill-off behavior is runner-generic,
# and skipping it saves ~9 min of mostly torch.compile time in CI. The
# no-silent-fallback assertion below is exercised by the V2 cases.
],
)
@pytest.mark.parametrize("enforce_eager", [True, False])
def test_kv_sharing_fast_prefill(
monkeypatch: pytest.MonkeyPatch,
kv_sharing_fast_prefill: bool,
enforce_eager: bool,
use_v2_model_runner: bool,
):
if not enforce_eager and current_platform.is_rocm():
# Relevant context: https://github.com/vllm-project/vllm/pull/29244
@@ -84,6 +95,7 @@ def test_kv_sharing_fast_prefill(
m.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
else:
m.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
m.setenv("VLLM_USE_V2_MODEL_RUNNER", "1" if use_v2_model_runner else "0")
prompts, answer, indices = prep_prompts(batch_size)
@@ -95,6 +107,8 @@ def test_kv_sharing_fast_prefill(
kv_sharing_fast_prefill=kv_sharing_fast_prefill,
attention_backend="TRITON_ATTN",
)
# Guard against a silent fallback to the other model runner.
assert llm.llm_engine.vllm_config.use_v2_model_runner == use_v2_model_runner
responses = llm.generate(prompts, sampling_params)
check_answers(
indices,
+64
View File
@@ -7,6 +7,8 @@ while keeping per-block content compact, so padding bytes at the end of each pag
never addressed by the logical view.
"""
from types import SimpleNamespace
import pytest
import torch
@@ -22,6 +24,7 @@ from vllm.v1.kv_cache_interface import (
MLAAttentionSpec,
compute_layout_strides,
)
from vllm.v1.worker.gpu import attn_utils
from vllm.v1.worker.gpu.attn_utils import (
get_attn_cg_support,
get_query_lens_mismatch_unsupported_backend,
@@ -118,6 +121,67 @@ def test_attention_checks_preserve_global_and_target_scoped_support():
)
def test_get_kv_sharing_fast_prefill_eligible_layers(monkeypatch: pytest.MonkeyPatch):
"""Fast prefill applies to the contiguous suffix of KV-sharing layers.
Draft-model layers register after the target model's and may share KV, so
they must not extend (or break) the target's eligible suffix.
"""
def check(
layer_names: list[str],
shared: dict[str, str],
draft_layer_names: set[str] | None = None,
) -> set[str]:
monkeypatch.setattr(
attn_utils,
"get_layers_from_vllm_config",
lambda *a, **k: {name: None for name in layer_names},
)
monkeypatch.setattr(attn_utils, "get_shared_kv_cache_layers", lambda *a: shared)
vllm_config = SimpleNamespace(
cache_config=SimpleNamespace(kv_sharing_fast_prefill=True)
)
return attn_utils.get_kv_sharing_fast_prefill_eligible_layers(
vllm_config, draft_layer_names
)
# No KV sharing: nothing is eligible.
assert check(["t0", "t1"], {}) == set()
# Trailing run of sharing layers (YOCO-style second half).
assert check(["t0", "t1", "t2", "t3"], {"t2": "t1", "t3": "t1"}) == {"t2", "t3"}
# A non-sharing layer after a sharing one breaks the suffix.
assert check(["t0", "t1", "t2", "t3"], {"t1": "t0", "t3": "t0"}) == {"t3"}
# KV-sharing draft layers at the end are collected without an exclusion...
assert check(
["t0", "t1", "t2", "t3", "d0", "d1"],
{"t2": "t1", "t3": "t1", "d0": "t1", "d1": "t1"},
) == {"t2", "t3", "d0", "d1"}
# ...so the runner excludes them: skipped, not collected, and they do not
# break the target's trailing run.
assert check(
["t0", "t1", "t2", "t3", "d0", "d1"],
{"t2": "t1", "t3": "t1", "d0": "t1", "d1": "t1"},
draft_layer_names={"d0", "d1"},
) == {"t2", "t3"}
# Feature flag off: nothing is eligible even with sharing layers.
monkeypatch.setattr(
attn_utils, "get_layers_from_vllm_config", lambda *a, **k: {"t0": None}
)
monkeypatch.setattr(
attn_utils, "get_shared_kv_cache_layers", lambda *a: {"t0": "t0"}
)
vllm_config = SimpleNamespace(
cache_config=SimpleNamespace(kv_sharing_fast_prefill=False)
)
assert attn_utils.get_kv_sharing_fast_prefill_eligible_layers(vllm_config) == set()
def test_reshape_padded_kv_cache_strides_by_padded_page():
num_blocks = 3
spec = FullAttentionSpec(
+1 -1
View File
@@ -84,7 +84,7 @@ def test_qsa_circular_group_uses_custom_slot_mapping(monkeypatch):
monkeypatch.setattr(
model_runner_module,
"init_attn_backend",
lambda *args: ([], attn_cg_support, [8, 262144]),
lambda *args, **kwargs: ([], attn_cg_support, [8, 262144]),
)
monkeypatch.setattr(
model_runner_module,