LLM Inference Workstation — 4x RTX6000 Blackwell Pro Max-Q (384GB VRAM) / Threadripper Pro 7985WX / WRX90E-SAGE SE / 512GB RAM / 1700W

Got deepseek-v4.1-flash running here through 0xSero/deepseek-v4.1-flash-4x-rtx-pro-6000 sglang’s runtime. Very impressed, here are the aggregate stats for my real traffic (same workload: cc, codex, agent evals) vs previous deployment of deepseek-v4-flash-0731:

Metric DSV4-FLASH-0731 DSV4.1-FLASH
Duration
Wall-clock observation span 531.70 h 47.55 h
Sum of request-activity windows 511.73 h 38.43 h
Observed launches 16 8 (1 running)
Completed non-health generations 256,325 10,871
Health-check generations 6,303 2,661
POST /v1/messages 175,668 8,651
POST /v1/chat/completions 62,436 1,768
POST /v1/responses 18,662 550
POST /v1/messages/count_tokens 2,356 86
Prompt size, tokens
Average full prompt 111,499 113,199
Median prompt 102,047 108,197
p95 prompt 264,293 251,338
Largest prompt 473,107 524,177
Average cached prompt tokens 107,688 89,025
Average uncached prompt tokens 3,811 24,174
Latency, seconds — p50 / p95 / p99
Queue / admission 0.64 / 2.08 / 27.14 0.84 / 78.98 / 204.46
TTFT, engine-side 1.21 / 7.09 / 42.00 1.38 / 107.15 / 232.27
Prefill 0.27 / 2.27 / 18.09 0.30 / 38.05 / 48.43
Decode 5.04 / 49.89 / 152.32 3.23 / 86.35 / 251.02
E2E, engine-side 6.47 / 58.60 / 161.35 5.37 / 186.02 / 335.03
Throughput, average
Per-request decode, tok/s 48.8 102.0
Batch generation, tok/s 137.4 191.9
Prefill input, tok/s 2,444 5,955
DSPARK(5) acceptance length / rate N/A — not enabled 3.71 / 0.542
Tokens
Input tokens 28,579,867,197 1,230,587,082
Output tokens 148,540,907 8,652,596
Of which reported reasoning 87,042,866 5,690,092
Average output per generation 580 796
Cache & health
Token-weighted cache hit 96.58% 78.64%
Zero-hit generations 11,578 (4.52%) 2,127 (19.57%)
HTTP 200 / 503 / 400, all routes 303,043 / 40 / 37 21,438 / 26 / 0
Other HTTP statuses 404: 54, 500: 10 0
HTTP 2xx rate, all routes 99.953% 99.879%
HTTP 2xx rate, inference routes only 99.982% 100.000%
Allocator OOM-warning log lines 0 1,470

OxSero’s DSV4.1-flash recipe is mostly as is, with exception to engram OFFLOAD_MODE=ram and sglang’s RAM HiCache enabled at 256GB (total offloaded at 468.8GiB) and swa caching patches per Retain reusable SWA request windows in RAM HiCache · ambientlight/sglang@28cedd2 · GitHub, old dsv4-flash-0731 per https://hub.docker.com/r/ambientlight/sglang-sm120-mxfp4?tag=2026.08.0-cu130-sm120a, launch params diff per:

Metric DSV4-FLASH-0731 DSV4.1-FLASH
Docker container dsv4 dsv41
Served model name deepseek-v4-flash deepseek-v4-flash
Tensor / expert parallelism TP4 / EP1 TP4 / EP4
Context limit, tokens 1,048,576 524,288
Static memory fraction 0.85 0.87
Maximum running requests 16 8
Chunked prefill size, tokens 4,096 2,048
Maximum prefill tokens (--max-prefill-tokens) 16,384 16,384
Prefill/decode interval 4 4
KV capacity sizing (--max-total-tokens) Automatic; unset Automatic; unset
Observed main KV capacity, logical tokens⁴ 5,677,312 3,367,936
Observed GPU SWA pool, slots per rank⁴ 567,552 26,112
KV cache dtype fp8_e4m3 fp8_e4m3
KV page size, tokens 256 256
Prefix-cache eviction policy LRU LRU + SWA boundary retention
RAM HiCache budget Disabled 256 GB host-total; FULL + SWA
HiCache write / I/O / layout N/A write_through / kernel / page_first
NVMe KV-cache tier Disabled Disabled
Engram offload mode N/A ram; separate NVMe row cache disabled
Speculative decoding Disabled DSPARK, block size 5
CUDA graph batch-size limit 16 8 (decode)
CUDA expandable segments Enabled Disabled
Attention backend dsv4 dsv4
MoE runner backend flashinfer_mxfp4 flashinfer_mxfp4
Reasoning / tool-call parser deepseek-v4 / deepseekv4 deepseek-v41 / deepseekv41
Default temperature / top-p 1.0 / 0.95 1.0 / 0.95
1 Like