[Misc][Nixl] Unify _logical_to_remote_kernel_block_ids (#48717)

Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
This commit is contained in:
Nicolò Lucchesi
2026-07-16 18:36:25 +08:00
committed by GitHub
parent a9531edfa6
commit a317bc5739
5 changed files with 41 additions and 64 deletions
@@ -86,7 +86,9 @@ def test_logical_to_kernel_block_ids_with_hma():
# Test conversion: FA + SW group
logical_block_ids = [[0, 1, 2], [3, 4]]
kernel_block_ids = worker._logical_to_kernel_block_ids(logical_block_ids)
kernel_block_ids = worker._logical_to_kernel_block_ids(
logical_block_ids, worker._physical_blocks_per_logical_kv_block
)
expected_kernel_block_ids = [[0, 1, 2, 3, 4, 5], [6, 7, 8, 9]]
assert kernel_block_ids == expected_kernel_block_ids, (
@@ -179,7 +181,7 @@ def test_read_blocks_for_req_expands_remote_ids(
"""_read_blocks_for_req must expand remote logical block IDs to kernel
block IDs when kernel block size != logical block size.
The hot path always calls _logical_to_remote_kernel_block_ids with
The hot path always calls _logical_to_kernel_block_ids with
remote_info.remote_physical_blocks_per_logical (model-agnostic).
"""
from unittest.mock import MagicMock
@@ -1140,7 +1142,7 @@ def test_derive_mamba_conv_split(
),
],
)
def test_logical_to_remote_kernel_block_ids(
def test_logical_to_kernel_block_ids_with_remote_ratio(
mamba_enabled,
swa_enabled,
local_physical_per_logical,
@@ -1148,7 +1150,7 @@ def test_logical_to_remote_kernel_block_ids(
logical_block_ids,
expected_kernel_block_ids,
):
"""Verify _logical_to_remote_kernel_block_ids uses the remote
"""Verify _logical_to_kernel_block_ids uses the remote
physical_per_logical for FA expansion, not the local one.
This was the root cause of silent accuracy corruption in Qwen3.5
@@ -1169,7 +1171,7 @@ def test_logical_to_remote_kernel_block_ids(
swa_enabled=swa_enabled,
)
result = worker._logical_to_remote_kernel_block_ids(
result = worker._logical_to_kernel_block_ids(
logical_block_ids,
remote_physical_per_logical,
)
@@ -487,7 +487,7 @@ class TestPushWriterStartLoadKv:
# path here.
w._send_heartbeats = lambda metadata: None
# Stub logical-to-kernel mapping used by reqs_to_recv.
w._logical_to_kernel_block_ids = lambda x: x
w._logical_to_kernel_block_ids = lambda x, ratio: x
meta = NixlConnectorMetadata()
meta.push_registrations = {
@@ -775,7 +775,7 @@ class TestPushWriterNegative:
"""Empty metadata must not wake the writer or enqueue anything."""
w = _StubWriterWorker.fresh()
w._send_heartbeats = lambda metadata: None
w._logical_to_kernel_block_ids = lambda x: x
w._logical_to_kernel_block_ids = lambda x, ratio: x
meta = NixlConnectorMetadata()
w.start_load_kv(meta)
@@ -944,7 +944,7 @@ class TestPushWriterMlaReplication:
rank_offset_factor=0,
)
}
w._logical_to_remote_kernel_block_ids = lambda block_ids, ratio: block_ids
w._logical_to_kernel_block_ids = lambda block_ids, ratio: block_ids
w.dst_xfer_side_handles = {engine_id: {r: 1000 + r for r in d_ranks}}
w.src_xfer_handles_by_block_size = {16: 2000}
w._remote_agents = {engine_id: {(0, r): f"agent-{r}" for r in d_ranks}}