Revert DSv4 eager workspace reuse (#52836)

Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Woosuk Kwon
2026-08-18 17:02:59 -07:00
committed by GitHub
co-authored by OpenAI Codex
parent a9f4afb66f
commit f1178f3a06
15 changed files with 30 additions and 354 deletions
-18
View File
@@ -20,7 +20,6 @@ import torch
from vllm import _custom_ops as ops
from vllm.models.deepseek_v4.common.ops import (
compute_global_topk_indices_and_lens,
dequantize_and_gather_k_cache,
quantize_and_insert_k_cache,
)
@@ -113,23 +112,6 @@ def test_get_dspark_swa_index_width(
assert get_dspark_swa_index_width(window_size, num_speculative_tokens) == expected
def test_compute_global_topk_reuses_output_buffers():
device = "cuda"
topk_indices = torch.tensor(
[[0, 3, -1], [1, 2, -1]], dtype=torch.int32, device=device
)
token_to_req = torch.tensor([0, 1], dtype=torch.int32, device=device)
block_table = torch.tensor([[5, 7], [11, 13]], dtype=torch.int32, device=device)
is_valid = torch.tensor([True, False], device=device)
args = (topk_indices, token_to_req, block_table, 2, is_valid)
expected = compute_global_topk_indices_and_lens(*args)
outputs = tuple(torch.empty_like(tensor) for tensor in expected)
actual = compute_global_topk_indices_and_lens(*args, output_buffers=outputs)
for result, output, reference in zip(actual, outputs, expected):
assert result.data_ptr() == output.data_ptr()
torch.testing.assert_close(result, reference)
def _ue8m0_reference(x: torch.Tensor, block_size: int, fp8_max: float):
"""PyTorch reference for UE8M0 FP8 quantization (per-block, power-of-2 scale).