| 2026-07-05 | Qwen3.6-27B | NVFP4 (native FP4) | 1× PRO 6000 / vastnode002 (+I2V) @300W | vLLM 0.24 + flashinfer 120f | 200K ctx, max-seqs 36, MTP n=2, util 0.54 | ~78-80 tok/s warm (measured) | ~297 tok/s aggregate (measured) | ~19.5 GB wt (shares card w/ Wan I2V) | sm120 FP4 crash-loop FIXED via flashinfer compute_120f (FLASHINFER_FORCE_SM=120f) — the AOT 120a cutlass FP4 GEMM was broken on Blackwell. n=3 MTP is 5× worse; autotune-on crashes. ~100 needs faster kernel (nightly staged). |
| 2026-06-15 | Qwen3.6-27B | FP8 | 1× PRO 6000 / vastnode002 (+I2V) | vLLM 0.22.1 | 200K ctx, max-seqs 32, MTP spec, util 0.58 | 72.3 tok/s (measured) | 363 @8 · 1558 @32 streams (measured) | 55.6 GB (shares card w/ Wan I2V) | Moved off 5000s→6000; ~1.8× single / ~1.5× batched vs the old 5000 (40/1044). Coexists w/ Wan I2V. |
| 2026-06-14 | Qwen3-Omni-30B-A3B Instruct | AWQ-int4 | A5000 GPU1 / r730a | vLLM 0.22.1 | 20K ctx · image≤2 · video 1×128f · max-seqs 4 | 133.8 tok/s (measured) | 426 tok/s @4 streams (107/strm, measured) | 19.2 GB wt / ~22.7 GB run | MoE 3B-active → fast decode; image:3+ OOMs at video profiling |
| 2026-06-14 | Qwen2.5-VL-7B | int8 | 7900 XTX GPU1 / vm905 | custom /vl-chat (ROCm) | ~200 tok gen | ~4 tok/s (measured) | batch TBD | ~9 GB | 50.7s for ~200 tok — SLOW for 7B int8; investigate ROCm path / per-call overhead |
| 2026-06-14 | LTX-2.3 22B (t2av) | fp8 self-quant | PRO 6000 / vastnode002 | LightX2V | ~10 s clip, 768p, +audio | ~33 s/clip (measured) | — (1 stream) | ~27 GB | No longer the pipeline bottleneck |
| 2026-06-14 | Qwen-Image-2512 (T2I) | bnb int8 + remote 7B TE | A5000 GPU0 / r730a | diffusers | 1024², 40 steps | 247.7 s/img (measured) | — (1 stream) | 20.9 GB peak | Re-measured (was 190); remote-TE round-trip to vm905 adds latency. Pipeline bottleneck. |
| 2026-06-23 | ACE-Step v1.5 turbo (music) | 2B DiT + 1.7B LM | 4070S Ti / 3930a | diffusers + vllm | 30s track, audio_duration 30 | ~12 s/track (measured) | batch up to 4 | ~10 GB | Tier-5 (12-16GB): max 8min/track w/ LM, 10min DiT-only. Apache, commercial-OK. Replaced VibeVoice. |
| 2026-06-14 | VibeVoice-7B (retired) | bnb int8 | 4070S Ti / 3930a | transformers SDPA | ~70-word narration | 0.53× realtime (measured) | batch TBD | ~13 GB | RETIRED 2026-06-23 (replaced by ACE-Step; Qwen handles voice). 46.8s gen for 24.8s audio — slower than realtime. |
| 2026-06-23 | Qwen-Image-Edit-2511 + 8-step Lightning | GGUF Q6_K + bf16 LoRA | 7900 XTX GPU0 / vm905 | diffusers GGUF (ROCm) + peft | 768², multi-image edit, 8-step cfg1.0 | ~49 s/edit warm (measured) | — serialized (lock) | ~17 GB | 8-step Lightning LoRA (lightx2v 2511) = 2.4× vs ~120s @20-step. peft req'd for LoRA-on-GGUF. Concurrent /edit serialized via lock (ROCm core-dumps on parallel pipe calls). Cold run +75s MIOpen compile. |
| 2026-06-15 | Wan2.2 T2V-A14B | NVFP4 | PRO 5000 ×2 / vastnode003 | LightX2V | 720p, 81 frames, 4-step distill | ~41 s/clip (measured) | 1 clip/worker; 2 workers ≈ 1.6× single-6000 | ~40 GB peak (8 GB headroom) | Moved 6000→5000s. 6000 was ~33s/clip same config. Server uses config res, ignores API params. |
| 2026-07-09 | dense-27b (Qwen3.6-27B) | NVFP4 + MTP n2 | 1× PRO 5000 (GPU1) / vastnode003 | vLLM nightly | 800 tok, temp0, max-model-len 150k, seqs 16, util 0.92 | ~94.7 tok/s (measured) | @8=510 · @16=970 · @24=828 tok/s agg (measured) | 44/49 GB | One PRO 5000 replica. Single-stream 94.7 = 82% of the 6000's ~115 (BW ratio). Agg still LINEAR 8→16 (510→970) = NOT saturated at seqs 16 — the ~1550 BW ceiling needs seqs~40, which 150k ctx won't allow on 48GB (KV 128KB/tok). @24 dips (seqs-16 admission cap). So 2×5000 @ this config ≈ ~1940 agg (parity w/ one 6000 @ high-conc), NOT the 1.5×; the 1.5× needs concurrency the big-context config forbids. 22W idle→check under load. |