diff --git a/vllm/models/glm5next/amd/ops/kpool_compress.py b/vllm/models/glm5next/amd/ops/kpool_compress.py index 7afc67e99e..64e6fd46d7 100644 --- a/vllm/models/glm5next/amd/ops/kpool_compress.py +++ b/vllm/models/glm5next/amd/ops/kpool_compress.py @@ -631,9 +631,11 @@ def kpool_decode_update_and_maybe_write_cache_batched( tail_kv_cache: paged tail cache ``[num_blocks, 2, pool_size, head_dim]`` bf16 (K at half 0, gate score at half 1). tail_slot_mapping: ``[num_requests, next_n]`` int32. - key / slot_score: ``[num_requests, next_n, head_dim]`` bf16. + key: ``[num_requests, next_n, head_dim]`` bf16. + slot_score: ``[num_requests, next_n, head_dim]`` bf16. ape: ``[pool_size, head_dim]`` fp32. - slot_mapping / positions: ``[num_requests, next_n]`` int32. + slot_mapping: ``[num_requests, next_n]`` int32. + positions: ``[num_requests, next_n]`` int32. """ num_requests, next_n = key.shape[0], key.shape[1] if num_requests == 0 or next_n == 0: diff --git a/vllm/models/glm5next/nvidia/__init__.py b/vllm/models/glm5next/nvidia/__init__.py new file mode 100644 index 0000000000..208f01a7cb --- /dev/null +++ b/vllm/models/glm5next/nvidia/__init__.py @@ -0,0 +1,2 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project diff --git a/vllm/models/glm5next/nvidia/ops/__init__.py b/vllm/models/glm5next/nvidia/ops/__init__.py new file mode 100644 index 0000000000..208f01a7cb --- /dev/null +++ b/vllm/models/glm5next/nvidia/ops/__init__.py @@ -0,0 +1,2 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project diff --git a/vllm/models/glm5next/nvidia/ops/kpool_compress.py b/vllm/models/glm5next/nvidia/ops/kpool_compress.py index 35b51c6b74..70b41021aa 100644 --- a/vllm/models/glm5next/nvidia/ops/kpool_compress.py +++ b/vllm/models/glm5next/nvidia/ops/kpool_compress.py @@ -637,9 +637,11 @@ def kpool_decode_update_and_maybe_write_cache_batched( tail_kv_cache: paged tail cache ``[num_blocks, 2, pool_size, head_dim]`` bf16 (K at half 0, gate score at half 1). tail_slot_mapping: ``[num_requests, next_n]`` int32. - key / slot_score: ``[num_requests, next_n, head_dim]`` bf16. + key: ``[num_requests, next_n, head_dim]`` bf16. + slot_score: ``[num_requests, next_n, head_dim]`` bf16. ape: ``[pool_size, head_dim]`` fp32. - slot_mapping / positions: ``[num_requests, next_n]`` int32. + slot_mapping: ``[num_requests, next_n]`` int32. + positions: ``[num_requests, next_n]`` int32. """ num_requests, next_n = key.shape[0], key.shape[1] if num_requests == 0 or next_n == 0: