Skip to content

[DSV4.1][HiCache] fix: wait for the layer transfer before reading low-ratio index-K #41345

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
hnyls2002 merged 3 commits into from
Sep 27, 2026
Merged

[DSV4.1][HiCache] fix: wait for the layer transfer before reading low-ratio index-K #41345

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
58df976
fix: wait for the layer transfer before reading DSV4.1 low-ratio index-K
alphabetc1 Sep 26, 2026
7ca6e0b
Merge branch 'main' into fix/dsv41-low-ratio-index-k-wait
alphabetc1 Sep 26, 2026
8f93168
Merge branch 'main' into fix/dsv41-low-ratio-index-k-wait
alphabetc1 Sep 26, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1867,6 +1867,7 @@ def get_low_ratio_index_k_dequant(
self, layer_id: int, slots: Optional[torch.Tensor] = None
) -> torch.Tensor:
"""Index-K rows at `slots` from the layer's latent source."""
self.wait_layer_transfer(layer_id)
compress_ratio, compress_layer_id, _ = self.layer_mapping[layer_id]
return self._indexer_pool(compress_ratio).get_index_k_dequant(
compress_layer_id, slots
Expand All @@ -1877,6 +1878,7 @@ def get_low_ratio_index_k_fp4(
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Packed fp4 index-K rows at `slots`: (payload int8 [n, 64], ue8m0 scales
packed int32 [n]), the input layout of quantize_fp4_indexer_tensor."""
self.wait_layer_transfer(layer_id)
compress_ratio, compress_layer_id, _ = self.layer_mapping[layer_id]
return self._indexer_pool(compress_ratio).get_index_k_fp4(
compress_layer_id, slots
Expand Down
6 changes: 6 additions & 0 deletions test/registered/unit/mem_cache/test_dsv4_compressed_pools.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -205,12 +205,18 @@ def test_indexer_access_uses_layer_ratio_and_waits_only_for_reads(self):
with patch.object(pool, "wait_layer_transfer") as wait:
trace.attach_mock(wait, "wait")
pool.get_index_k_fp4_payload_buffer(2)
pool.get_low_ratio_index_k_dequant(2, "slots")
pool.get_low_ratio_index_k_fp4(2, "slots")
pool.set_index_k_fp4(2, "loc", "cache")
self.assertEqual(
trace.mock_calls,
[
unittest.mock.call.wait(2),
unittest.mock.call.indexer.get_index_k_fp4_payload_buffer(1),
unittest.mock.call.wait(2),
unittest.mock.call.indexer.get_index_k_dequant(1, "slots"),
unittest.mock.call.wait(2),
unittest.mock.call.indexer.get_index_k_fp4(1, "slots"),
unittest.mock.call.indexer.set_index_fp4(1, "loc", "cache"),
],
)
Expand Down
Loading