[MRV2] Miscellaneous code cleanup (#57980)
This commit is contained in:
@@ -166,9 +166,7 @@ def test_build_draft_attn_metadata_recomputes_dcp_local_seq_lens():
|
||||
out[num_reqs:].zero_()
|
||||
return out
|
||||
|
||||
with patch.object(
|
||||
base_speculator, "maybe_prepare_dcp_local_seq_lens", fake_prepare
|
||||
):
|
||||
with patch.object(base_speculator, "prepare_dcp_local_seq_lens", fake_prepare):
|
||||
captured = _run_build(
|
||||
fake,
|
||||
num_reqs=3,
|
||||
|
||||
@@ -81,7 +81,7 @@ def test_make_dummy_padding_controls_moe_routing(monkeypatch, is_padding: bool):
|
||||
assert bool((topk_ids >= 0).all()) is not is_padding
|
||||
|
||||
|
||||
def test_maybe_prepare_dcp_local_seq_lens_uses_shared_buffer(monkeypatch):
|
||||
def test_prepare_dcp_local_seq_lens_uses_shared_buffer(monkeypatch):
|
||||
"""The batch must view the caller-owned buffer, sliced to padded length.
|
||||
|
||||
Runtime (and capture) paths all funnel through this helper; attention
|
||||
@@ -115,7 +115,7 @@ def test_maybe_prepare_dcp_local_seq_lens_uses_shared_buffer(monkeypatch):
|
||||
return fake_kernel
|
||||
|
||||
monkeypatch.setattr(cp_utils, "_dcp_local_seq_lens_kernel", FakeKernel())
|
||||
batch.dcp_local_seq_lens = cp_utils.maybe_prepare_dcp_local_seq_lens(
|
||||
batch.dcp_local_seq_lens = cp_utils.prepare_dcp_local_seq_lens(
|
||||
buffers.dcp_local_seq_lens,
|
||||
batch.seq_lens,
|
||||
batch.num_reqs,
|
||||
@@ -130,37 +130,10 @@ def test_maybe_prepare_dcp_local_seq_lens_uses_shared_buffer(monkeypatch):
|
||||
assert batch.dcp_local_seq_lens.tolist() == [1, 2, 0, 0]
|
||||
|
||||
|
||||
def test_maybe_prepare_dcp_local_seq_lens_clears_stale_metadata(monkeypatch):
|
||||
"""dcp_size == 1 resets the field to None instead of leaving a leftover.
|
||||
|
||||
DCP is toggled per deployment, and batches are recycled; a value from an
|
||||
earlier DCP batch would otherwise be consumed as if it were current.
|
||||
"""
|
||||
buffers = InputBuffers(max_num_reqs=2, max_num_tokens=2, device=torch.device("cpu"))
|
||||
batch = InputBatch.make_dummy(1, 1, buffers)
|
||||
batch.dcp_local_seq_lens = buffers.dcp_local_seq_lens[:1]
|
||||
|
||||
def fail_if_called(*args, **kwargs):
|
||||
raise AssertionError("kernel must not run with dcp_size == 1")
|
||||
|
||||
monkeypatch.setattr(cp_utils, "_dcp_local_seq_lens_kernel", fail_if_called)
|
||||
batch.dcp_local_seq_lens = cp_utils.maybe_prepare_dcp_local_seq_lens(
|
||||
buffers.dcp_local_seq_lens,
|
||||
batch.seq_lens,
|
||||
batch.num_reqs,
|
||||
1,
|
||||
0,
|
||||
1,
|
||||
num_reqs_padded=batch.num_reqs_after_padding,
|
||||
)
|
||||
|
||||
assert batch.dcp_local_seq_lens is None
|
||||
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="triton kernel needs CUDA")
|
||||
@pytest.mark.parametrize("dcp_size", [2, 4])
|
||||
@pytest.mark.parametrize("cp_interleave", [1, 16])
|
||||
def test_maybe_prepare_dcp_local_seq_lens_matches_reference(
|
||||
def test_prepare_dcp_local_seq_lens_matches_reference(
|
||||
dcp_size: int, cp_interleave: int
|
||||
):
|
||||
"""Every rank's local lengths must equal the reference torch formula."""
|
||||
@@ -172,7 +145,7 @@ def test_maybe_prepare_dcp_local_seq_lens_matches_reference(
|
||||
|
||||
for dcp_rank in range(dcp_size):
|
||||
buffers.dcp_local_seq_lens.fill_(-1)
|
||||
batch.dcp_local_seq_lens = cp_utils.maybe_prepare_dcp_local_seq_lens(
|
||||
batch.dcp_local_seq_lens = cp_utils.prepare_dcp_local_seq_lens(
|
||||
buffers.dcp_local_seq_lens,
|
||||
batch.seq_lens,
|
||||
batch.num_reqs,
|
||||
|
||||
@@ -2,7 +2,6 @@
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
from dataclasses import replace
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import MagicMock
|
||||
|
||||
import numpy as np
|
||||
import pytest
|
||||
@@ -170,38 +169,44 @@ def test_graph_padding_cannot_be_smaller_than_largest_pcp_rank(monkeypatch):
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="needs GPU kernels")
|
||||
@pytest.mark.parametrize(
|
||||
("cg_mode", "num_reqs", "expected_tokens", "expected_reqs"),
|
||||
[
|
||||
(CUDAGraphMode.NONE, 4, None, None),
|
||||
(CUDAGraphMode.PIECEWISE, None, 8, None),
|
||||
(CUDAGraphMode.NONE, 4, 2, 2),
|
||||
(CUDAGraphMode.PIECEWISE, None, 8, 2),
|
||||
(CUDAGraphMode.FULL, 4, 8, 4),
|
||||
],
|
||||
)
|
||||
def test_partition_padding_is_derived_from_batch_descriptor(
|
||||
cg_mode, num_reqs, expected_tokens, expected_reqs
|
||||
):
|
||||
manager = MagicMock()
|
||||
input_batch = MagicMock()
|
||||
manager.partition_batch.return_value = input_batch
|
||||
batch_desc = BatchExecutionDescriptor(
|
||||
cg_mode=cg_mode,
|
||||
num_tokens=8,
|
||||
num_reqs=num_reqs,
|
||||
"""partition_batch derives graph padding from the batch descriptor: no
|
||||
padding without graphs, token padding only for PIECEWISE, and request
|
||||
padding on top for FULL."""
|
||||
device = torch.device("cuda:0")
|
||||
manager = PCPManager(
|
||||
pcp_world_size=2,
|
||||
pcp_rank=0,
|
||||
device=device,
|
||||
max_num_reqs=4,
|
||||
max_num_tokens=8,
|
||||
)
|
||||
global_batch = _make_global_decode_batch(
|
||||
[16, 24], InputBuffers(4, 8, device), device
|
||||
)
|
||||
|
||||
result = pcp_manager_module.maybe_partition_pcp_batch(
|
||||
manager,
|
||||
input_batch,
|
||||
batch_desc,
|
||||
local_batch = manager.partition_batch(
|
||||
global_batch,
|
||||
BatchExecutionDescriptor(
|
||||
cg_mode=cg_mode,
|
||||
num_tokens=8,
|
||||
num_reqs=num_reqs,
|
||||
),
|
||||
)
|
||||
|
||||
assert result is input_batch
|
||||
manager.partition_batch.assert_called_once_with(
|
||||
input_batch,
|
||||
padded_num_tokens=expected_tokens,
|
||||
padded_num_reqs=expected_reqs,
|
||||
)
|
||||
assert local_batch.num_tokens_after_padding == expected_tokens
|
||||
assert local_batch.num_reqs_after_padding == expected_reqs
|
||||
|
||||
|
||||
def test_capture_uses_pcp_persistent_inputs():
|
||||
@@ -472,8 +477,11 @@ def test_partition_defers_dcp_metadata_to_post_partition_batch():
|
||||
|
||||
local_batch = manager.partition_batch(
|
||||
global_batch,
|
||||
padded_num_tokens=4,
|
||||
padded_num_reqs=4,
|
||||
BatchExecutionDescriptor(
|
||||
cg_mode=CUDAGraphMode.FULL,
|
||||
num_tokens=4,
|
||||
num_reqs=4,
|
||||
),
|
||||
)
|
||||
|
||||
assert local_batch.dcp_local_seq_lens is None
|
||||
@@ -489,7 +497,7 @@ def test_partition_defers_dcp_metadata_to_post_partition_batch():
|
||||
|
||||
# What execute_model does next: derive DCP metadata from the final batch
|
||||
# on the PCP-owned buffers.
|
||||
local_batch.dcp_local_seq_lens = gpu_cp_utils.maybe_prepare_dcp_local_seq_lens(
|
||||
local_batch.dcp_local_seq_lens = gpu_cp_utils.prepare_dcp_local_seq_lens(
|
||||
manager.input_buffers.dcp_local_seq_lens,
|
||||
local_batch.seq_lens,
|
||||
local_batch.num_reqs,
|
||||
|
||||
@@ -756,7 +756,7 @@ def test_microbatches_recompute_dcp_lens_from_truncated_seq_lens(dcp_rank: int):
|
||||
[1, 1, 10, 1, 1], [64, 96, 512, 32, 48]
|
||||
)
|
||||
# execute_model has populated the merged batch's DCP metadata already.
|
||||
input_batch.dcp_local_seq_lens = gpu_cp_utils.maybe_prepare_dcp_local_seq_lens(
|
||||
input_batch.dcp_local_seq_lens = gpu_cp_utils.prepare_dcp_local_seq_lens(
|
||||
buffers.dcp_local_seq_lens,
|
||||
input_batch.seq_lens,
|
||||
input_batch.num_reqs,
|
||||
|
||||
Reference in New Issue
Block a user