Recipe
glm52-nvfp4-aqlm-dgxspark-vllm-tp3
glm52-nvfp4-aqlm-dgxspark-vllm-tp3MiaAI-Lab GLM-5.2 hybrid NVFP4+AQLM vision profile across three DGX Sparks with TP3, DCP1, MTP-3, and full CUDA graphs
Record
- Status
- candidate
- Source
- mialabs
- Engine
- vllm
- Engine version
- jarrelscy glm52-sm120 fork
- Graph
- full
- Accelerators
- 3
- Tensor parallel
- 3
- Context tokens
- 40,000
- Max concurrency
- 1
- chat
- yes
- reasoning
- yes
- tools
- yes
- vision
- yes
Hugging Face model card
Identity
https://huggingface.co/jarrelscy/GLM-5.2-NVFP4-AQLM-hybrid- Repository
- jarrelscy/GLM-5.2-NVFP4-AQLM-hybrid
- Status
- known
- Link type
- Exact Hub repository
Public Hugging Face repository confirmed by the Hub API.
Candidate: useful compatibility or speed evidence. The registry does not offer Run until promotion requirements are met.
Script configuration
vllm
No container · candidate · script
Candidate evidence — not a Run contract
Environment
| Variable | Value |
|---|---|
DCP_SIZE | 1 |
ENABLE_MTP | 1 |
GPU_MEM_UTIL | 0.895 |
HF_REVISION | 53e0082eedebd806b63e19779c47905937d768ca |
KV_CACHE_DTYPE | nvfp4_ds_mla |
KV_CACHE_MEMORY_BYTES | 11811160064 |
MAX_MODEL_LEN | 348160 |
MAX_NUM_BATCHED_TOKENS | 4096 |
MAX_NUM_SEQS | 1 |
MTP_SPEC_TOKENS | 3 |
TP_SIZE | 3 |
Measured speed
| Concurrency | Context | Prefill | Decode | TTFT ms | Status | Sweep |
|---|---|---|---|---|---|---|
| 1 | — | — | 21 | — | historical | glm52-nvfp4-aqlm-dgxspark-vllm-tp3-sweep |
| 1 | 40,000 | — | 13.7 | — | historical | glm52-nvfp4-aqlm-dgxspark-vllm-tp3-sweep |
Remaining fields
Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.
- hardware count
- 3
- hardware id
- dgx-spark-gb10-128gb
- id
- glm52-nvfp4-aqlm-dgxspark-vllm-tp3
- model instance id
- jarrelscy-glm-5-2-nvfp4-aqlm-hybrid--nvfp4-hot-experts-aqlm-2-bit-cold-experts
- recipe source
- mialabs
- schema version
- local-ai-registry/v1
- speed sweep ids
- glm52-nvfp4-aqlm-dgxspark-vllm-tp3-sweep
- status
- candidate
capabilities
- chat
- Yes
- reasoning
- Yes
- tools
- Yes
- vision
- Yes
engine
- graph mode
- full
- name
- vllm
- version
- jarrelscy glm52-sm120 fork
serving
- max concurrency
- 1
- max context tokens
- 40,000
- tensor parallel
- 3
Provenance & metadata (3)
facts
- serving.kv cache tokens · provenance · captured at
- 2026-08-27T06:04:13.773Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
metadata
provenance
- captured at
- 2026-08-27T06:04:13.773Z
sources
| captured at | kind | url |
|---|---|---|
| 2026-08-27T06:04:13.773Z | normalized-recipe | github.com/0xSero/local-ai-registry ↗ |