[KV Cache] GLM-5.3-Flash: use the generic packed KV layout; CircularBufferSpec tail with speculative ring slots (#57169)

Signed-off-by: Yifan Qiao <yifanqiao@inferact.ai>
Signed-off-by: Dakai An <dakaian108@gmail.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
This commit is contained in:
Dakai An
2026-10-09 15:48:37 -04:00
committed by GitHub
co-authored by Yifan Qiao OpenAI Codex OpenAI Codex Lucas Wilkinson Claude Opus 5.5 Lucas Wilkinson
parent 3766a67a32
commit 02e55d2af3
83 changed files with 1280 additions and 1986 deletions
+3 -5
View File
@@ -55,8 +55,9 @@ def resolve_hisparse_specs(
if len(block_sizes) != 1:
raise ValueError("HiSparse requires one scheduler block size.")
backends = [attn_layers[name].get_attn_backend() for name in mla_specs]
specs = list(mla_specs.values())
try:
block_size = select_common_block_size(block_sizes.pop(), backends)
block_size = select_common_block_size(block_sizes.pop(), backends, specs)
except ValueError as error:
raise ValueError(
"HiSparse requires a GPU block size supported by every sparse "
@@ -104,16 +105,13 @@ def allocate_hisparse_kv_caches(
if isinstance(host_spec, UniformTypeKVCacheSpecs)
else host_spec
)
kernel_block_size = kernel_block_sizes[host_group_id]
if isinstance(spec, MLAAttentionSpec) and spec.storage_block_size is not None:
kernel_block_size = spec.storage_block_size
views = create_kv_cache_views(
backing,
spec,
kv_cache_config.num_blocks_of(tensor),
layout,
tensor,
kernel_block_size=kernel_block_size,
kernel_block_size=kernel_block_sizes[host_group_id],
)
kv_caches.update(zip(tensor.layers, views))
return kv_caches
+2 -2
View File
@@ -333,7 +333,7 @@ def get_hisparse_kv_cache_config(
device_groups = hisparse_layout.device_groups
layout = vllm_config.cache_config.get_resolved_kv_cache_layout()
validate_kv_cache_layout(layout, device_groups)
bytes_per_block = _get_kv_cache_bytes_per_block(device_groups)
bytes_per_block = _get_kv_cache_bytes_per_block(device_groups, layout)
num_blocks = may_override_num_blocks(
vllm_config, available_memory // bytes_per_block
)
@@ -345,7 +345,7 @@ def get_hisparse_kv_cache_config(
host_groups = [hisparse_layout.source_group]
host_layout = KVCacheLayout.LBNHC
validate_kv_cache_layout(host_layout, host_groups)
host_bytes_per_block = _get_kv_cache_bytes_per_block(host_groups)
host_bytes_per_block = _get_kv_cache_bytes_per_block(host_groups, host_layout)
host_size = host_bytes_per_block * hisparse_layout.host_num_blocks
kv_cache_tensors[:0] = _build_hisparse_kv_cache_tensors(
host_groups,