Speed sweep

glm-5-3-flash-nvfp4-dgx-spark-gb10-128gb-vllm-tp4-sweep

glm-5-3-flash-nvfp4-dgx-spark-gb10-128gb-vllm-tp4-sweep

Record

Recipe
glm-5-3-flash-nvfp4-dgx-spark-gb10-128gb-vllm-tp4
Measured
2026-09-01
Accepted
2026-09-01
Points
1

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatus
18,1821,764.9112.14,636.4accepted-sustained

Remaining fields

Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.

accepted at
2026-09-01
id
glm-5-3-flash-nvfp4-dgx-spark-gb10-128gb-vllm-tp4-sweep
measured at
2026-09-01
recipe id
glm-5-3-flash-nvfp4-dgx-spark-gb10-128gb-vllm-tp4
schema version
local-ai-registry/v1

metrics

base memory bytes
358800470016
base memory context tokens
0
concurrency
1
decode8k context tokens
8,182
decode8k tps
112.118
decode mode
C1 structured JSON-schema matched-window with natural stop
inference engine version
0.1.dev20051+g487ecf187
latest point at
2026-09-01
max context tokens
524,288
max prompt tokens
8,183
memory8k bytes
359068327936
memory8k context tokens
8,182
peak generation tps
117.992
peak memory bytes
359068327936
peak prompt tps
1,771.873
point count
1

rows

all stream windowbase unified memory gibcache stateconcurrencycontent classcontext tokensdecode tok sdecode tok s per stream
first-emitted-token-through-last-emitted-token334.159cold-unique-cache-salt1structured-json-schema8,182112.118112.118
Provenance & metadata (1)

source

kind
live-runtime-capture
paths
docs/notes/glm-5-3-flash-nvfp4-dgx-spark-vllm-tp4-20260901.json, docs/notes/glm-5-3-flash-nvfp4-dgx-spark-vllm-tp4-compose.yml, registry/asset/glm-5-3-flash-multimodal-chat-template.jinja, registry/asset/glm-5-3-flash-vllm-sparse-attn-indexer-kpool.py