Speed sweep

deepseek-v4-flash-fp8-rtxpro6000-sglang-tp4-sweep

deepseek-v4-flash-fp8-rtxpro6000-sglang-tp4-sweep

Record

Recipe
deepseek-v4-flash-fp8-rtxpro6000-sglang-tp4
Measured
2026-08-24
Accepted
2026-08-24
Points
7

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatus
18,192701.637.611,337historical
116,3841,2503812,770historical
132,7681,273.135.525,101historical
165,5361,145.130.355,850historical
1131,07294624.1135,271historical
1196,000793.819.3246,863historical
1300,000646.415.1464,072historical

Remaining fields

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

accepted at
2026-08-24
id
deepseek-v4-flash-fp8-rtxpro6000-sglang-tp4-sweep
measured at
2026-08-24
recipe id
deepseek-v4-flash-fp8-rtxpro6000-sglang-tp4
schema version
local-ai-registry/v1

metrics

concurrency
1
latest point at
2026-08-24
max context tokens
300,000
peak generation tps
37.99
peak prompt tps
1,273.1
point count
7

rows

concurrencycontext tokensdecode tok sdecode tok s per streamoutput tokenspeak vram gbprefill tok ssamples
18,19237.5937.59UnknownUnknown701.61
116,38437.9937.99UnknownUnknown1,2501
132,76835.535.5UnknownUnknown1,273.11
165,53630.3330.33UnknownUnknown1,145.11
1131,07224.0824.08UnknownUnknown9461
1196,00019.3319.33UnknownUnknown793.81
1300,00015.0915.09UnknownUnknown646.41
Provenance & metadata (1)

source

commit
c2eac5a9b2b457881d69b1164d909e8beab9286e
paths
README.md, DEEPSEEK-V4-FLASH.md, scripts/launch_dsv4_flash_sm120.sh, scripts/build_in_sglang_docker.sh, deepseek_v4_kernel/_patch.py