GLM-5.3#

A large Mixture-of-Experts model using Dynamic Sparse Attention (DSA), continuing the GLM-5.1 / GLM-5.2 line (architecture GlmMoeDsaForCausalLM, 78 layers, index_topk 2048). Unlike DeepSeek-V4.1-Flash, this checkpoint presents a single KV cache group to the connector, so no per-group tuning is needed.

Validated models#

zai-org/GLM-5.3 (8 GPUs)

vLLM

Engine documentation: GLM-5.3 in vLLM supported models (architecture GlmMoeDsaForCausalLM).

Status: Validated with LMCache.

Start the LMCache MP server:

lmcache server \
    --chunk-size 256 \
    --separate-object-groups \
    --l1-size-gb 100 \
    --eviction-policy LRU

Start vLLM with the LMCache MP connector (8 GPUs):

vllm serve zai-org/GLM-5.3 \
    --tensor-parallel-size 8 \
    --enable-prefix-caching \
    --kv-transfer-config \
    '{"kv_connector":"LMCacheMPConnector", "kv_role":"kv_both"}'

Chunk size. GLM-5.3 reports a single KV cache group with tokens_per_block = 64, so --chunk-size must be a multiple of 64. Both 256 and 1024 were validated; prefer the smaller value, because only whole chunks are stored and the remainder is recomputed on every request. For a 5040-token prompt, --chunk-size 1024 caches 4096 tokens (81%) while 256 caches 4864 (97%) – a measured external-hit-rate difference of 40.6% vs 48.3% on that prompt. Choose a chunk size well below your typical prompt length.

The connector logs the geometry it resolved at startup as group_tokens_per_block=[...], scheduler_block_size=... – check that line if you are unsure what a given checkpoint reports.

Adjust --tensor-parallel-size to match your hardware. For the generic LMCache + vLLM wiring (ports, remote hosts), see Quickstart. GLM’s serving extras (--tool-call-parser, --reasoning-parser) are orthogonal to LMCache; see the GLM-5.1/5.2 recipe and the vLLM GLM recipe for those.

SGLang

Status: Not validated with LMCache.

TRT-LLM

Status: Not validated with LMCache.

CacheBlend support#

Not validated.

Compression support#

Method	Status	Notes
CacheGen	Not validated

MTP (speculative decoding) support#

GLM-5.3 ships a native multi-token-prediction head (num_nextn_predict_layers: 1), usable via vLLM’s --speculative-config '{"method":"mtp","num_speculative_tokens":1}'. Status: Not yet validated with LMCache.

Caveats#

Single KV cache group. Unlike the multi-group sparse-attention models in this section, the validated zai-org/GLM-5.3 checkpoint reports one group (group_tokens_per_block=[64]). A -Flash variant of this architecture additionally keeps a per-request kpool-tail scratch group (KpoolTailSpec), which LMCache skips automatically; see Hybrid Attention Models. That variant is not covered by this validation.
Cached pages are byte-opaque views, so cache entries must not be shared across engines with different attention backends or kernel block sizes.
The LMCache validation covers text generation with greedy decoding; the cold and cache-served runs reproduced each other exactly.