[Model] add GLM-5.3-Flash support (#53906)

This commit is contained in:
Jiangyun Zhu
2026-09-03 09:40:35 -07:00
committed by GitHub
parent da8ec28268
commit 98ed0856f3
94 changed files with 17927 additions and 266 deletions
+15
View File
@@ -763,6 +763,17 @@ class Platform:
if cache_config.mamba_page_size_padded is not None:
cache_config.mamba_page_size_padded = shared_page
@classmethod
def _get_indexer_block_alignment(cls, vllm_config: "VllmConfig") -> int | None:
"""Extra ``block_size`` multiple a sparse indexer needs, else ``None``.
The CUDA kpool paged-MQA indexer virtually splits each storage block
into pool pages, so ``block_size`` must be a multiple of
``index_kpool * min(PAGED_MQA_PAGE_SIZES)`` — implemented in the CUDA
platform override. Other platforms impose no extra constraint.
"""
return None
@classmethod
def _align_hybrid_block_size(
cls,
@@ -805,6 +816,7 @@ class Platform:
num_kv_heads=model_config.get_num_kv_heads(parallel_config),
head_size=model_config.get_head_size(),
dtype=kv_cache_dtype,
cache_dtype_str=cache_config.cache_dtype,
kv_quant_mode=kv_quant_mode,
).page_size_bytes
elif cache_config.cache_dtype.startswith("turboquant_"):
@@ -912,6 +924,9 @@ class Platform:
mamba_page_size,
kernel_block_alignment_size * attn_page_size_1_token,
)
indexer_align = cls._get_indexer_block_alignment(vllm_config)
if indexer_align:
attn_block_size = indexer_align * cdiv(attn_block_size, indexer_align)
if cache_config.block_size < attn_block_size:
cache_config.block_size = attn_block_size