[Bugfix][Docs] Package glm5next nvidia subtree and fix its docstrings (#55214)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Harry Mellor committed
Sep 4, 2026 at 05:25 UTC
78300cdabffbaaffe7625ed34412ec2aec75535c
4 files changed
+12
-4
vllm/models/glm5next/amd/ops/kpool_compress.py
+4
-2
@@ -631,9 +631,11 @@ def kpool_decode_update_and_maybe_write_cache_batched(
631
tail_kv_cache: paged tail cache ``[num_blocks, 2, pool_size, head_dim]``
632
bf16 (K at half 0, gate score at half 1).
633
tail_slot_mapping: ``[num_requests, next_n]`` int32.
634
- key / slot_score: ``[num_requests, next_n, head_dim]`` bf16.
634
+ key: ``[num_requests, next_n, head_dim]`` bf16.
635
+ slot_score: ``[num_requests, next_n, head_dim]`` bf16.
636
ape: ``[pool_size, head_dim]`` fp32.
636
- slot_mapping / positions: ``[num_requests, next_n]`` int32.
637
+ slot_mapping: ``[num_requests, next_n]`` int32.
638
+ positions: ``[num_requests, next_n]`` int32.
639
"""
640
num_requests, next_n = key.shape[0], key.shape[1]
641
if num_requests == 0 or next_n == 0:
vllm/models/glm5next/nvidia/__init__.py
new
+2
@@ -0,0 +1,2 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
vllm/models/glm5next/nvidia/ops/__init__.py
new
+2
@@ -0,0 +1,2 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
vllm/models/glm5next/nvidia/ops/kpool_compress.py
+4
-2
@@ -637,9 +637,11 @@ def kpool_decode_update_and_maybe_write_cache_batched(
637
tail_kv_cache: paged tail cache ``[num_blocks, 2, pool_size, head_dim]``
638
bf16 (K at half 0, gate score at half 1).
639
tail_slot_mapping: ``[num_requests, next_n]`` int32.
640
- key / slot_score: ``[num_requests, next_n, head_dim]`` bf16.
640
+ key: ``[num_requests, next_n, head_dim]`` bf16.
641
+ slot_score: ``[num_requests, next_n, head_dim]`` bf16.
642
ape: ``[pool_size, head_dim]`` fp32.
642
- slot_mapping / positions: ``[num_requests, next_n]`` int32.
643
+ slot_mapping: ``[num_requests, next_n]`` int32.
644
+ positions: ``[num_requests, next_n]`` int32.
645
"""
646
num_requests, next_n = key.shape[0], key.shape[1]
647
if num_requests == 0 or next_n == 0: