DeepSeek V4 · Flash 180B

attached MTP depth × TP

Exact experimental K160 artifact · PIECEWISE · FP8 KV. Historical TP4 rows use EP4 and retain their exact differing runtime identities; rates are not transferred across them. The approximately 103.1 GB on-disk target does not fit fully resident at TP1 or TP2 on 32/64 GiB; CPU offload is outside this view. Fixed: revision=deepseek-v4-flash-180b-community · artifact_id=deepseek-v4-flash-180b-k160-7c360e1 · variant=experimental uniform-K160 FP8/MXFP4 target · quantization=experimental uniform-K160 FP8/MXFP4 target · graph_mode=PIECEWISE · kv=fp8.

• ✓ Measured4 cards, target-only TP4·target-only43.77 tok/sevidence
• ✓ Measured4 cards, MTP1 TP4·MTP163.85 tok/sevidence
• ■ Closed4 cards, MTP2 TP4·MTP2 — M=3 capture and 10/10 exact pre-hang rows passed, but the second draft accepted only 0.5–2.2% and realistic traffic stopped before a valid suite completed.no valid rate
• ■ Closed4 cards, MTP4 TP4·MTP4 — The evidence-based post-MTP3 stop decision closed four drafted tokens after the third proposal accepted at most 3.2%. No MTP4 throughput exists.not launched
• × Unsupported1 card, target-only TP1·target-only — The approximately 103.1 GB on-disk K160 target cannot fit fully resident on one 32 GiB B70; CPU offload is outside this view.weights do not fit
• × Unsupported2 cards, target-only TP2·target-only — The approximately 103.1 GB on-disk K160 target cannot fit fully resident on two 32 GiB B70s; CPU offload is outside this view.weights do not fit
• × Unsupported1 card, MTP1 TP1·MTP1 — The fully resident target already exceeds one B70 before the attached MTP layer is considered.weights do not fit
• × Unsupported2 cards, MTP1 TP2·MTP1 — The fully resident target already exceeds two B70s before the attached MTP layer is considered.weights do not fit
• × Unsupported1 card, MTP2 TP1·MTP2 — The fully resident target already exceeds one B70; repeated-layer drafting cannot change target capacity.weights do not fit
• × Unsupported2 cards, MTP2 TP2·MTP2 — The fully resident target already exceeds two B70s; repeated-layer drafting cannot change target capacity.weights do not fit
• × Unsupported1 card, MTP3 TP1·MTP3 — The fully resident target already exceeds one B70; repeated-layer drafting cannot change target capacity.weights do not fit
• × Unsupported2 cards, MTP3 TP2·MTP3 — The fully resident target already exceeds two B70s; repeated-layer drafting cannot change target capacity.weights do not fit
• × Unsupported1 card, MTP4 TP1·MTP4 — The fully resident target already exceeds one B70; repeated-layer drafting cannot change target capacity.weights do not fit
• × Unsupported2 cards, MTP4 TP2·MTP4 — The fully resident target already exceeds two B70s; repeated-layer drafting cannot change target capacity.weights do not fit
• ⚠ Quarantined4 cards, MTP3 TP4·MTP3 — Only 8/10 rows reached the timing window; the third draft accepted 0–3.2%, so the 46.247281 diagnostic is not a valid endpoint result. Observed: D46.247 · incomplete.evidence

Codes in the rows: D = decode tok/s · P = prefill tok/s · T = ms to first token · AR = share of drafted tokens accepted · A = combined tok/s.

speculation × TP

Exact experimental K160 target and record runtime; missing cells have no inherited result. Fixed: revision=deepseek-v4-flash-180b-community · artifact_id=deepseek-v4-flash-180b-k160-7c360e1 · variant=experimental uniform-K160 FP8/MXFP4 target · runtime=vLLM XPU record stack 264c7f2/3131567/48fda4f · graph_mode=PIECEWISE.

• ✓ Measured4 cards, DSpark7 TP4·DSpark780.82 tok/sguide · evidence
• × Unsupported1 card, DSpark7 TP1·DSpark7 — The approximately 103.1 GB on-disk K160 target cannot fit fully resident on one 32 GiB B70, before the DSpark draft pack is considered; CPU offload is outside this view.weights do not fit
• × Unsupported2 cards, DSpark7 TP2·DSpark7 — The approximately 103.1 GB on-disk K160 target cannot fit fully resident on two 32 GiB B70s, before the DSpark draft pack is considered; CPU offload is outside this view.weights do not fit