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).
@@ -257,18 +257,8 @@ def test_q_path_matches_reference(num_tokens: int, n_heads: int, padded_heads: i
num_blocks, bs, HEAD_BYTES, dtype=torch.uint8, device=device
).view(num_blocks, -1)
slot_mapping = torch.full((num_tokens,), -1, dtype=torch.int64, device=device)
q_out = torch.empty(num_tokens, padded_heads, HEAD_DIM, dtype=dtype, device=device)
torch.ops._C.fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert_out(
q,
kv,
q_out,
k_cache,
slot_mapping,
positions,
cos_sin_cache,
padded_heads,
eps,
bs,
q_out = _call_fused(
q, padded_heads, kv, k_cache, slot_mapping, positions, cos_sin_cache, eps, bs
)
torch.testing.assert_close(q_out[:, :n_heads], q_ref, rtol=1e-2, atol=1e-2)
@@ -150,23 +150,6 @@ def test_fused_indexer_q_rope_quant_matches_unfused(
q_quant_ref, weights_ref = _reference(
positions, q, cos_sin_cache, weights, softmax_scale, head_scale, use_fp4
)
output_buffers: tuple[torch.Tensor, ...] | None = None
OUTPUT_BUFFER_TEST_NUM_TOKENS = 7
if num_tokens == OUTPUT_BUFFER_TEST_NUM_TOKENS and cache_dtype == torch.float32:
if use_fp4:
q_ref, q_scale_ref = q_quant_ref
output_buffers = (
torch.empty_like(q_ref),
torch.empty_like(q_scale_ref)
.view(torch.uint8)
.reshape(num_tokens, N_HEAD, -1),
torch.empty_like(weights_ref),
)
else:
output_buffers = (
torch.empty_like(q_quant_ref),
torch.empty_like(weights_ref),
)
# use_cutedsl=False: force the triton path even when cutedsl is installed
# by patching the dispatcher's has_cutedsl() binding to return False.
cutedsl_patch = (
@@ -186,17 +169,8 @@ def test_fused_indexer_q_rope_quant_matches_unfused(
softmax_scale,
head_scale,
use_fp4,
output_buffers=output_buffers,
)
if output_buffers is not None:
if use_fp4:
assert q_quant_fused[0].data_ptr() == output_buffers[0].data_ptr()
assert q_quant_fused[1].data_ptr() == output_buffers[1].data_ptr()
else:
assert q_quant_fused.data_ptr() == output_buffers[0].data_ptr()
assert weights_fused.data_ptr() == output_buffers[-1].data_ptr()
if use_fp4:
q_quant_ref, q_scale_ref = q_quant_ref
q_quant_fused, q_scale_fused = q_quant_fused