[DecodeBenchConnector] Fix HMA cache-group mapping (#54647)
Signed-off-by: Jeff Ma <jeffjma@umich.edu> Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
co-authored by
OpenAI Codex
parent
40b2f62061
commit
30dd1a7954
@@ -7,6 +7,8 @@ Tests the functionality of the DecodeBenchConnector which fills KV cache
|
||||
with dummy values for decode performance benchmarking.
|
||||
"""
|
||||
|
||||
from unittest.mock import MagicMock
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
@@ -22,6 +24,11 @@ from vllm.forward_context import ForwardContext
|
||||
from vllm.utils.hashing import sha256
|
||||
from vllm.v1.core.kv_cache_utils import get_request_block_hasher, init_none_hash
|
||||
from vllm.v1.core.sched.scheduler import Scheduler
|
||||
from vllm.v1.kv_cache_interface import (
|
||||
FullAttentionSpec,
|
||||
KVCacheConfig,
|
||||
KVCacheGroupSpec,
|
||||
)
|
||||
from vllm.v1.request import Request
|
||||
|
||||
from .utils import (
|
||||
@@ -65,11 +72,10 @@ class DecodeBenchTestRunner:
|
||||
# Using simplified shape for testing
|
||||
num_heads = 4
|
||||
head_dim = 64
|
||||
layer_names = self.scheduler.kv_cache_config.kv_cache_groups[0].layer_names
|
||||
self.kv_caches = {
|
||||
f"layer_{i}": torch.zeros(
|
||||
num_gpu_blocks, 2, num_heads, block_size, head_dim
|
||||
)
|
||||
for i in range(2) # 2 layers for testing
|
||||
layer_name: torch.zeros(num_gpu_blocks, 2, num_heads, block_size, head_dim)
|
||||
for layer_name in layer_names
|
||||
}
|
||||
|
||||
# Register KV caches with worker connector
|
||||
@@ -179,6 +185,97 @@ def test_decode_bench_connector_basic():
|
||||
assert torch.allclose(block_data, torch.tensor(0.015))
|
||||
|
||||
|
||||
def test_decode_bench_connector_fills_each_hma_group():
|
||||
"""Each cache group is filled using its own block IDs."""
|
||||
block_size = 16
|
||||
num_gpu_blocks = 8
|
||||
vllm_config = create_vllm_config(
|
||||
block_size=block_size,
|
||||
max_num_batched_tokens=1000,
|
||||
kv_connector="DecodeBenchConnector",
|
||||
)
|
||||
kv_cache_spec = FullAttentionSpec(
|
||||
block_size=block_size,
|
||||
num_kv_heads=1,
|
||||
head_size=1,
|
||||
dtype=torch.float32,
|
||||
)
|
||||
kv_cache_config = KVCacheConfig(
|
||||
num_blocks=num_gpu_blocks,
|
||||
kv_cache_tensors=[],
|
||||
kv_cache_groups=[
|
||||
KVCacheGroupSpec(["group_0_layer"], kv_cache_spec),
|
||||
KVCacheGroupSpec(["group_1_layer"], kv_cache_spec),
|
||||
],
|
||||
)
|
||||
connector = DecodeBenchConnector(
|
||||
vllm_config,
|
||||
KVConnectorRole.WORKER,
|
||||
kv_cache_config,
|
||||
)
|
||||
group_0_cache = torch.zeros(num_gpu_blocks, 2)
|
||||
group_1_cache = torch.zeros(num_gpu_blocks, 2)
|
||||
connector.register_kv_caches(
|
||||
{
|
||||
"group_0_layer": group_0_cache,
|
||||
"group_1_layer": group_1_cache,
|
||||
}
|
||||
)
|
||||
connector.bind_connector_metadata(
|
||||
DecodeBenchConnectorMetadata(reqs_to_fill={"request": (([1], [5]), block_size)})
|
||||
)
|
||||
|
||||
connector.start_load_kv(
|
||||
ForwardContext(no_compile_layers={}, attn_metadata={}, slot_mapping={})
|
||||
)
|
||||
|
||||
assert torch.allclose(group_0_cache[1], torch.tensor(0.015))
|
||||
assert torch.count_nonzero(group_0_cache[5]) == 0
|
||||
assert torch.count_nonzero(group_1_cache[1]) == 0
|
||||
assert torch.allclose(group_1_cache[5], torch.tensor(0.015))
|
||||
|
||||
|
||||
def test_decode_bench_connector_uses_per_group_block_sizes():
|
||||
"""Scheduler selects only the blocks covering each group's token span."""
|
||||
vllm_config = create_vllm_config(
|
||||
block_size=16,
|
||||
max_num_batched_tokens=1000,
|
||||
kv_connector="DecodeBenchConnector",
|
||||
)
|
||||
kv_cache_groups = [
|
||||
KVCacheGroupSpec(
|
||||
[f"group_{group_idx}_layer"],
|
||||
FullAttentionSpec(
|
||||
block_size=block_size,
|
||||
num_kv_heads=1,
|
||||
head_size=1,
|
||||
dtype=torch.float32,
|
||||
),
|
||||
)
|
||||
for group_idx, block_size in enumerate((16, 32))
|
||||
]
|
||||
connector = DecodeBenchConnector(
|
||||
vllm_config,
|
||||
KVConnectorRole.SCHEDULER,
|
||||
KVCacheConfig(
|
||||
num_blocks=8,
|
||||
kv_cache_tensors=[],
|
||||
kv_cache_groups=kv_cache_groups,
|
||||
),
|
||||
)
|
||||
request = MagicMock(request_id="request")
|
||||
blocks = MagicMock()
|
||||
blocks.get_block_ids.return_value = ([0, 1], [2, 3])
|
||||
|
||||
connector.update_state_after_alloc(request, blocks, num_external_tokens=17)
|
||||
metadata = connector.build_connector_meta(MagicMock())
|
||||
|
||||
assert isinstance(metadata, DecodeBenchConnectorMetadata)
|
||||
block_ids_per_group, num_tokens = metadata.reqs_to_fill["request"]
|
||||
assert block_ids_per_group == ([0, 1], [2])
|
||||
assert num_tokens == 17
|
||||
|
||||
|
||||
def test_decode_bench_connector_no_refill():
|
||||
"""Test that DecodeBenchConnector only fills once per request."""
|
||||
block_size = 16
|
||||
|
||||
Reference in New Issue
Block a user