[Core] Add MRV2 virtual-batch PCP for MLA (#46570)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Lucas Wilkinson
2026-07-19 02:53:15 +00:00
committed by GitHub
co-authored by Codex OpenAI Codex Matthew Bonanni
parent 9243e0124e
commit b6ff8a2f50
39 changed files with 1204 additions and 196 deletions
@@ -0,0 +1,16 @@
model_name: "nvidia/GLM-5.2-NVFP4"
accuracy_threshold: 0.90
num_questions: 1319
num_fewshot: 5
max_concurrency: 100
server_args: >-
--enforce-eager
--max-model-len 4096
--safetensors-load-strategy prefetch
--moe-backend flashinfer_cutlass
--prefill-context-parallel-size 4
--enable-expert-parallel
--kv-cache-dtype fp8
env:
VLLM_LOGGING_LEVEL: "DEBUG"
VLLM_USE_V2_MODEL_RUNNER: "1"
@@ -0,0 +1,17 @@
model_name: "nvidia/GLM-5.2-NVFP4"
accuracy_threshold: 0.90
num_questions: 1319
num_fewshot: 5
max_concurrency: 100
server_args: >-
--enforce-eager
--max-model-len 4096
--safetensors-load-strategy prefetch
--moe-backend flashinfer_cutlass
--tensor-parallel-size 2
--prefill-context-parallel-size 2
--enable-expert-parallel
--kv-cache-dtype fp8
env:
VLLM_LOGGING_LEVEL: "DEBUG"
VLLM_USE_V2_MODEL_RUNNER: "1"
+2
View File
@@ -0,0 +1,2 @@
GLM-5.2-NVFP4-TP2-PCP2-EP.yaml
GLM-5.2-NVFP4-TP1-PCP4-EP.yaml
+15 -1
View File
@@ -217,6 +217,7 @@ def evaluate_gsm8k(
seed: int | None = 42,
request_timeout_seconds: float = 600,
gen_prefix: str = "",
max_concurrency: int | None = None,
) -> dict[str, float | int]:
"""
Evaluate GSM8K accuracy using vLLM serve endpoint.
@@ -261,7 +262,14 @@ def evaluate_gsm8k(
return answer, tokens
timeout = aiohttp.ClientTimeout(total=request_timeout_seconds)
async with aiohttp.ClientSession(timeout=timeout) as session:
connector = (
aiohttp.TCPConnector(limit=max_concurrency)
if max_concurrency is not None
else None
)
async with aiohttp.ClientSession(
timeout=timeout, connector=connector
) as session:
tasks = [get_answer(session, i) for i in range(num_questions)]
await tqdm.gather(*tasks, desc="Evaluating")
@@ -343,6 +351,11 @@ def main() -> None:
parser.add_argument(
"--seed", type=int, default=42, help="Random seed for reproducibility"
)
parser.add_argument(
"--max-concurrency",
type=int,
help="Maximum number of concurrent requests",
)
parser.add_argument("--save-results", type=str, help="Save results to JSON file")
args = parser.parse_args()
@@ -355,6 +368,7 @@ def main() -> None:
port=args.port,
temperature=args.temperature,
seed=args.seed,
max_concurrency=args.max_concurrency,
)
# Print results to terminal
@@ -59,6 +59,7 @@ def run_gsm8k_eval(eval_config: dict, server_url: str) -> dict:
seed=eval_config.get("seed", 42),
request_timeout_seconds=request_timeout_seconds,
gen_prefix=eval_config.get("gen_prefix", ""),
max_concurrency=eval_config.get("max_concurrency"),
)
return results
@@ -160,6 +160,8 @@ def apply_split_decodes_and_prefills(
decode_threshold: int,
require_uniform: bool,
padded_num_tokens: int | None = None,
is_prefilling: list[bool] | None = None,
treat_short_extends_as_decodes: bool = True,
):
"""Helper function to apply split_decodes_and_prefills and return
the results."""
@@ -173,11 +175,14 @@ def apply_split_decodes_and_prefills(
if padded_num_tokens is not None:
common_metadata.num_actual_tokens = padded_num_tokens
if is_prefilling is not None:
common_metadata.is_prefilling = torch.tensor(is_prefilling)
return split_decodes_and_prefills(
common_metadata,
decode_threshold=decode_threshold,
require_uniform=require_uniform,
treat_short_extends_as_decodes=treat_short_extends_as_decodes,
)
@@ -236,6 +241,17 @@ def test_split_decodes_and_prefills_uniform_all_ones():
assert num_prefill_tokens == 0
def test_split_decodes_and_prefills_uniform_short_extend():
result = apply_split_decodes_and_prefills(
[1, 1],
decode_threshold=1,
require_uniform=True,
is_prefilling=[False, True],
treat_short_extends_as_decodes=False,
)
assert result == (1, 1, 1, 1)
def test_split_decodes_and_prefills_uniform_all_short_decodes():
query_lens = [2, 2, 1, 3, 2, 1, 2]
num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens = (
@@ -1114,6 +1114,13 @@ def test_sparse_backend_prefill_correctness(
@pytest.mark.parametrize(
"seq_lens,query_lens,workspace_size,max_logits_bytes,expected",
[
(
torch.tensor([0]),
torch.tensor([0]),
100,
1000,
[],
),
# Logits constraint triggers split (M*N exceeds budget)
# req0: M=10, N=100 -> 1000 elems (4000 bytes) - fits in 5000
# req1: adding M=10, N=100 -> new_M=20, new_N=200 -> 4000 elems > 1250
+4 -4
View File
@@ -385,7 +385,7 @@ def test_tiering_spec_aligns_row_size():
assert spec.num_blocks == cpu_bytes_to_use // alignment
def test_offloading_spec_resolves_prefill_context_parallel_block_sizes():
def test_offloading_spec_kv_sharding_ignores_prefill_context_parallel():
config = _make_layout_vllm_config(
cpu_bytes_to_use=65536,
extra_config={"block_size": 64},
@@ -394,9 +394,9 @@ def test_offloading_spec_resolves_prefill_context_parallel_block_sizes():
spec = _create_spec(config, _make_kv_cache_config())
assert spec.tokens_per_block == (32,)
assert spec.tokens_per_hash == 32
assert spec.blocks_per_chunk == 2
assert spec.tokens_per_block == (16,)
assert spec.tokens_per_hash == 16
assert spec.blocks_per_chunk == 4
def test_offloading_config_preserves_data_parallel_index():
+16 -24
View File
@@ -1551,9 +1551,8 @@ def _make_cp_scheduler(
num_gpu_blocks: int = 16,
lazy: bool = False,
) -> SchedulerFixture:
"""Build a SimpleCPUOffloadScheduler with CP-scaled virtual block size."""
cp_world_size = dcp_world_size * pcp_world_size
virtual_block_size = BLOCK_SIZE * cp_world_size
"""Build a SimpleCPUOffloadScheduler with DCP-scaled block size."""
virtual_block_size = BLOCK_SIZE * dcp_world_size
kv_cache_config = _make_kv_cache_config(num_gpu_blocks)
vllm_config = _make_cp_vllm_config(dcp_world_size, pcp_world_size)
@@ -1645,17 +1644,14 @@ def _allocate_cp_gpu_blocks(
],
)
def test_cp_block_size_scaling(dcp_world_size: int, pcp_world_size: int) -> None:
"""Verify that the scheduler's block_size and cp_world_size are correctly
scaled when context parallelism is enabled."""
"""Verify block size scaling follows DCP ownership."""
fix = _make_cp_scheduler(
dcp_world_size=dcp_world_size, pcp_world_size=pcp_world_size
)
sched = fix.scheduler
expected_cp = dcp_world_size * pcp_world_size
assert sched.cp_world_size == expected_cp
assert sched.block_size == BLOCK_SIZE * expected_cp
assert sched.fa_block_size == BLOCK_SIZE * expected_cp
assert sched.cp_world_size == dcp_world_size
assert sched.block_size == BLOCK_SIZE * dcp_world_size
# ---------------------------------------------------------------------------
@@ -1682,8 +1678,7 @@ def test_cp_eager_store_and_load_roundtrip(
lazy=False,
)
sched = fix.scheduler
cp = dcp_world_size * pcp_world_size
vbs = BLOCK_SIZE * cp
vbs = BLOCK_SIZE * dcp_world_size
num_blocks = 2
req = _make_cp_request(num_blocks, vbs)
@@ -1749,18 +1744,16 @@ def test_cp_eager_store_and_load_roundtrip(
def test_cp_effective_block_size_store_and_load(
dcp_world_size: int, pcp_world_size: int
) -> None:
"""Verify ready_blocks_g (store) and n_take_g (load) use the effective
(physical * cp) block size, not the per-rank physical size."""
"""Verify store/load physical block counts scale with DCP, not PCP."""
fix = _make_cp_scheduler(
dcp_world_size=dcp_world_size, pcp_world_size=pcp_world_size
)
sched = fix.scheduler
gpu_pool = fix.gpu_block_pool
cp = dcp_world_size * pcp_world_size
vbs = BLOCK_SIZE * cp
vbs = BLOCK_SIZE * dcp_world_size
expected_blocks = 1
# Store: allocate 2 blocks, confirm only 1. Without the fix,
# ready_blocks_g = vbs / BLOCK_SIZE = 2, storing both blocks.
# Store one DCP-scaled logical block worth of tokens.
req = _make_cp_request(num_blocks=2, virtual_block_size=vbs)
gpu_blocks = _allocate_cp_gpu_blocks(gpu_pool, req, 2, vbs)
kv = KVCacheBlocks(blocks=(gpu_blocks,))
@@ -1772,12 +1765,11 @@ def test_cp_effective_block_size_store_and_load(
new_reqs={req.request_id: kv.get_block_ids()},
)
)
assert len(m1.store_gpu_blocks) == 1
assert len(m1.store_cpu_blocks) == 1
assert len(m1.store_gpu_blocks) == expected_blocks
assert len(m1.store_cpu_blocks) == expected_blocks
simulate_store_completion(sched, m1.store_event)
# Load: store 2 blocks from a second request, accept only 1 as external.
# Without the fix, n_take_g = vbs / BLOCK_SIZE = 2, loading both.
# Load one DCP-scaled logical block from a two-block CPU hit.
req2 = _make_cp_request(num_blocks=2, virtual_block_size=vbs)
kv2 = KVCacheBlocks(blocks=(_allocate_cp_gpu_blocks(gpu_pool, req2, 2, vbs),))
req2.num_computed_tokens = 2 * vbs
@@ -1801,7 +1793,7 @@ def test_cp_effective_block_size_store_and_load(
hit, _ = sched.get_num_new_matched_tokens(req3, num_computed_tokens=0)
assert hit == 2 * vbs
kv3 = KVCacheBlocks(blocks=(gpu_pool.get_new_blocks(2),))
kv3 = KVCacheBlocks(blocks=(gpu_pool.get_new_blocks(expected_blocks),))
sched.update_state_after_alloc(req3, kv3, num_external_tokens=vbs)
m3 = sched.build_connector_meta(
make_scheduler_output(
@@ -1810,8 +1802,8 @@ def test_cp_effective_block_size_store_and_load(
)
)
assert m3.load_event >= 0
assert len(m3.load_gpu_blocks) == 1
assert len(m3.load_cpu_blocks) == 1
assert len(m3.load_gpu_blocks) == expected_blocks
assert len(m3.load_cpu_blocks) == expected_blocks
assert m3.load_gpu_blocks == [kv3.get_block_ids()[0][0]]