Recipe
qwen3-8-27b-nvfp4-dgx-spark-gb10-128gb-sglang-tp1-ahj9l2q3
qwen3-8-27b-nvfp4-dgx-spark-gb10-128gb-sglang-tp1-ahj9l2q3Observed LocalMaxxing result. It is evidence for compatibility, not an executable launch contract.
Record
- Status
- candidate
- Source
- localmaxxing
- Engine
- sglang
- Engine version
- 0.0.0.dev1+g5f55db35e
- Accelerators
- 1
- Tensor parallel
- 1
- Context tokens
- 131,072
- Max concurrency
- 1
- chat
- unknown
- reasoning
- unknown
- tools
- unknown
- vision
- unknown
Hugging Face model card
Identity
https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4- Repository
- unsloth/Qwen3.8-27B-NVFP4
- 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
sglang
Evidence only · candidate · reference
Candidate evidence — not a Run contract
No tokenized launch fields. This is measured or documented compatibility, not a Docker launch.
Measured speed
| Concurrency | Context | Prefill | Decode | TTFT ms | Status | Sweep |
|---|---|---|---|---|---|---|
| 1 | 131,072 | — | 11 | 195.6 | observed | qwen3-8-27b-nvfp4-dgx-spark-gb10-128gb-sglang-tp1-ahj9l2q3-sweep |
Remaining fields
Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.
- hardware count
- 1
- hardware id
- dgx-spark-gb10-128gb
- id
- qwen3-8-27b-nvfp4-dgx-spark-gb10-128gb-sglang-tp1-ahj9l2q3
- model instance id
- unsloth-qwen3-8-27b-nvfp4--nvfp4
- recipe source
- localmaxxing
- schema version
- local-ai-registry/v1
- speed sweep ids
- qwen3-8-27b-nvfp4-dgx-spark-gb10-128gb-sglang-tp1-ahj9l2q3-sweep
- status
- candidate
capabilities
draft launch
- accelerator backend
- nvidia
- arguments
- python3, -m, sglang.launch_server, --model-path, unsloth/Qwen3.8-27B-NVFP4, --tp, 1, --host, 0.0.0.0, --port, 30000, --context-length, 131072
- container port
- 30,000
- host port
- 30,000
- image
- lmsysorg/sglang:dev-cu13@sha256:6cd4635214f279e0a43019f88e3120d407567640a58aa7dcc0085e3d91402cc4
- ipc
- host
- kind
- docker
- shm size
- 16g
- synthesized · generated at
- 2026-08-31T22:12:17Z
- synthesized · image provenance
- gemma-4-12b-it-nvfp4-rtxpro6000-sglang-tp1
- synthesized · template
- sglang-launch-server-v1
mounts
| read only | target |
|---|---|
| No | /root/.cache/huggingface |
engine
- name
- sglang
- version
- 0.0.0.dev1+g5f55db35e
serving
- max concurrency
- 1
- max context tokens
- 131,072
- tensor parallel
- 1
Provenance & metadata (3)
facts
- capabilities.chat · provenance · captured at
- 2026-08-27T06:04:13.773Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
- capabilities.reasoning · provenance · captured at
- 2026-08-27T06:04:13.773Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
- capabilities.tools · provenance · captured at
- 2026-08-27T06:04:13.773Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
- capabilities.vision · provenance · captured at
- 2026-08-27T06:04:13.773Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
- engine.graph mode · provenance · captured at
- 2026-08-27T06:04:13.773Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
- 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
- localmaxxing · hardware label
- GB10 Grace Blackwell
- localmaxxing · notes
- Single-stream decode, batch 1, one concurrent request. n=9 generations (3 prompt kinds x 3 repeats), 512 output tokens each, streaming, temperature 0.6 / top_p 0.95 / top_k 20. Context length 131072 is the CONFIGURED window, not the prompt depth; prompts here are short. mem-fraction-static 0.80, kv-cache-dtype fp8_e4m3, attention backend flashinfer, tp 1, max-running-requests 32. Idle box, one model loaded at a time, and all 6 cells of this battery ran back to back in one session on one machine. SPECULATIVE DECODING OFF, no draft model loaded. This is the no-draft baseline arm of a DFlash2 comparison, so it is deliberately low and should not be read as this stack's practical speed. Per-workload decode: code 10.97, reasoning 10.97, prose 10.99 tok/s; the headline is the mean of all 9. Spread across workloads is under 0.03 tok/s, which is what a decode path with no draft looks like. Quantization detail: Unsloth Dynamic V3.0 compressed-tensors mixed precision, MLP gate/up/down NVFP4 W4A4 group 16, attention and layers 56-63 MLP FP8, lm_head FP8 per-channel, MTP head BF16 unquantised. Served from a local copy, so the Hugging Face revision was not recorded; the checkpoint is pinned instead by its config.json sha256 1b3c71868d1299e52df6fc907deb202d5132b1ef0f72aae0ef6d15185dd53a5c. TTFT is the mean of the same 9 runs.
- localmaxxing · revision unpinned
- Yes
- localmaxxing · run id
- cmta4vbq501coli01ahj9l2q3
provenance
- captured at
- 2026-08-27T06:04:13.773Z
sources
| captured at | kind | url |
|---|---|---|
| 2026-08-27T06:04:13.773Z | normalized-recipe | www.localmaxxing.com/en/runs/cmta4vbq501coli01ahj9l2q3 ↗ |