Repository navigation
GLM-4.7-Flash does not appear to activate IndexCache on the SGLang patch path #5
Description
4.7-Flash` config uses:
architectures = ["Glm4MoeLiteForCausalLM"]model_type = "glm4_moe_lite"
and does not appear to be recognized by is_deepseek_nsa(). It also lacks explicit NSA/DSA fields such as index_topk,
index_n_heads, and index_head_dim.
Local check:
architectures = ['Glm4MoeLiteForCausalLM']
model_type = 'glm4_moe_lite'
index_topk_freq = 4
index_topk = None
is_deepseek_nsa = False
Benchmarks on H100 show effectively identical baseline vs IndexCache results, suggesting IndexCache is not actually active for
the public GLM-4.7-Flash checkpoint.
Could you clarify whether the public checkpoint is expected to work with the current patch, and if so, what the correct
supported setup is?
Activity