Recipe
glm-5-2-ud-iq4-xs-rtx-pro-6000-blackwell-96gb-llama-cpp-tp4
glm-5-2-ud-iq4-xs-rtx-pro-6000-blackwell-96gb-llama-cpp-tp4Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.
Record
- Status
- candidate
- Source
- localmaxxing
- Engine
- llama.cpp
- Engine version
- 0eca4d4
- Accelerators
- 4
- Tensor parallel
- 4
- Context tokens
- 65,536
- Max concurrency
- 1
- chat
- unknown
- reasoning
- unknown
- tools
- unknown
- vision
- unknown
Hugging Face model card
Identity
https://huggingface.co/unsloth/GLM-5.2-GGUF- Repository
- unsloth/GLM-5.2-GGUF
- 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.
Observed configuration
llama.cpp
Evidence only · candidate · reference
Candidate evidence — not a Run contract
Observed source tokens
Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.
llama-server--modelGLM-5.2-UD-IQ4_XS-00001-of-00009.gguf--host0.0.0.0--port8001--aliasGLM-5.2-UD-IQ4_XS--ctx-size65536--n-gpu-layers999--parallel1--flash-attnon--jinja--metrics
| Flag | Value |
|---|---|
--model | GLM-5.2-UD-IQ4_XS-00001-of-00009.gguf |
--host | 0.0.0.0 |
--port | 8001 |
--alias | GLM-5.2-UD-IQ4_XS |
--ctx-size | 65536 |
--n-gpu-layers | 999 |
--parallel | 1 |
--flash-attn | on |
Environment
| Variable | Value |
|---|---|
CUDA_VISIBLE_DEVICES | 0,1,2,3 |
LD_LIBRARY_PATH | ~/code/ggml-org/llama.cpp/build/bin |
Measured speed
| Concurrency | Context | Prefill | Decode | TTFT ms | Status | Sweep |
|---|---|---|---|---|---|---|
| 1 | 65,536 | 561.2 | 39.8 | 3,971.3 | observed | glm-5-2-ud-iq4-xs-rtx-pro-6000-blackwell-96gb-llama-cpp-tp4-sweep |
Remaining fields
Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.
- hardware count
- 4
- hardware id
- rtx-pro-6000-blackwell-96gb
- id
- glm-5-2-ud-iq4-xs-rtx-pro-6000-blackwell-96gb-llama-cpp-tp4
- model instance id
- unsloth-glm-5-2-gguf--ud-iq4-xs
- recipe source
- localmaxxing
- schema version
- local-ai-registry/v1
- speed sweep ids
- glm-5-2-ud-iq4-xs-rtx-pro-6000-blackwell-96gb-llama-cpp-tp4-sweep
- status
- candidate
capabilities
engine
- name
- llama.cpp
- version
- 0eca4d4
serving
- max concurrency
- 1
- max context tokens
- 65,536
- tensor parallel
- 4
Provenance & metadata (3)
facts
- capabilities.chat · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
- capabilities.reasoning · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
- capabilities.tools · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
- capabilities.vision · provenance · captured at
- 2026-08-30T09:10:02Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
- engine.graph mode · provenance · captured at
- 2026-08-30T09:10:02Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
- serving.kv cache tokens · provenance · captured at
- 2026-08-30T09:10:02Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
metadata
- localmaxxing · backend
- cuda
- localmaxxing · hardware label
- RTX PRO 6000 Blackwell
- localmaxxing · notes
- GLM-5.2-504B via Unsloth UD-IQ4_XS on 4x RTX PRO 6000 Blackwell Max-Q, llama.cpp source-build (0eca4d4), CUDA 13.1 sm_120. Companion run to my earlier 131072-ctx submission — this one is at --ctx-size 65536 to see how much VRAM+speed the smaller KV reservation buys. Answer: essentially nothing on speed (39.7 -> 39.7 tok/s output, 562 -> 561 prefill server-side) and only ~6.6 GB less peak VRAM. At UD-IQ4_XS the model weights dominate; KV cache for 131k tokens is only ~10 GB total, so halving it is not the lever people might expect. Same bench methodology as my other run: 2 warmup, 3 measurement, greedy decode, unique per-run prompt to bust prefix caching, server-side print_timing for tokSPrefill/tokSOut, client TTFT.
- localmaxxing · observed command
- CUDA_VISIBLE_DEVICES=0,1,2,3 LD_LIBRARY_PATH=~/code/ggml-org/llama.cpp/build/bin llama-server --model GLM-5.2-UD-IQ4_XS-00001-of-00009.gguf --host 0.0.0.0 --port 8001 --alias GLM-5.2-UD-IQ4_XS --ctx-size 65536 --n-gpu-layers 999 --parallel 1 --flash-attn on --jinja --metrics
- localmaxxing · run id
- cmr1ipwec01eeld01nn27xd9d
- localmaxxing · tokenized · arguments
- llama-server, --model, GLM-5.2-UD-IQ4_XS-00001-of-00009.gguf, --host, 0.0.0.0, --port, 8001, --alias, GLM-5.2-UD-IQ4_XS, --ctx-size, 65536, --n-gpu-layers, 999, --parallel, 1, --flash-attn, on, --jinja, --metrics
- localmaxxing · tokenized · environment · CUDA VISIBLE DEVICES
- 0,1,2,3
- localmaxxing · tokenized · environment · LD LIBRARY PATH
- ~/code/ggml-org/llama.cpp/build/bin
- localmaxxing · tokenized · fidelity
- faithful
provenance
- captured at
- 2026-08-30T09:10:02Z
sources
| captured at | kind | url |
|---|---|---|
| 2026-08-30T09:10:02Z | normalized-recipe | www.localmaxxing.com/en/runs/cmr1ipwec01eeld01nn27xd9d ↗ |