GlmMoeDsaForCausalLM#

A large Mixture-of-Experts model using Dynamic Sparse Attention (DSA), shared by the GLM-5.2 series. Like DeepSeek-V4-Flash, the sparse-attention path splits the model’s layers into more than one KV cache group; the LMCacheMPConnector stores and retrieves each group in its own block size, so KV reuse works without extra flags.

Validated models#

zai-org/GLM-5.2-FP8 (8 GPUs)

vLLM

Engine documentation: GLM-5.2 in vLLM supported models (architecture GlmMoeDsaForCausalLM). See also the vLLM GLM-5.2 recipe.

Status: Validated with LMCache (vLLM 0.23.0 + LMCache 0.4.7).

Start the LMCache MP server:

lmcache server \
    --port 6555 \
    --max-workers 8 \
    --l1-size-gb 100 \
    --eviction-policy LRU \
    --chunk-size 1024

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

vllm serve zai-org/GLM-5.2-FP8 \
    --tensor-parallel-size 8 \
    --tool-call-parser glm47 \
    --enable-auto-tool-choice \
    --reasoning-parser glm45 \
    --no-enable-prefix-caching \
    --kv-transfer-config \
    '{"kv_connector":"LMCacheMPConnector","kv_role":"kv_both","kv_connector_extra_config":{"lmcache.mp.port":6555}}'

--tool-call-parser glm47, --enable-auto-tool-choice, and --reasoning-parser glm45 are GLM-5.2’s serving requirements (see the vLLM recipe). --no-enable-prefix-caching routes all KV reuse through LMCache rather than vLLM’s in-engine prefix cache. The server’s --port 6555 must match lmcache.mp.port in the connector config; --max-workers is set to the tensor-parallel size. Adjust --tensor-parallel-size to match your hardware. For the generic LMCache + vLLM wiring (ports, remote hosts), see Quickstart.

If there are any issues with vLLM setup, please refer to the vLLM Recipes for more details.

SGLang

Status: Not validated with LMCache.

TRT-LLM

Status: Supported. See Quickstart for TRT-LLM + LMCache setup.

CacheBlend support#

Not validated.

Compression support#

Method	Status	Notes
CacheGen	Not validated

Caveats#

Dynamic Sparse Attention KV groups. GLM-5.2’s DSA path splits the model’s layers into more than one KV cache group with different block geometries. LMCache stores and retrieves each group in its own block size; no extra flags are required beyond the launch commands above.