b70-optimization-lab

DeepSeek V4 Flash REAP/XPU On Four B70s

This is the preserved experiment packet for the paused DeepSeek V4 Flash uniform-K160 lane on four Arc Pro B70 32 GB GPUs. The frontier closed on 2026-07-21 at a verified 80.820052 tok/s high. Start from the closed result packet, standalone repro, and closeout. The older staged plan and handoffs below remain historical research context.

Read in order:

Frozen Identity

official source: deepseek-ai/DeepSeek-V4-Flash
source revision: 60d8d70770c6776ff598c94bb586a859a38244f1
first runnable smoke candidate: uniform-K160 0xSero/DeepSeek-V4-Flash-180B revision 7c360e1cd4a5168099dbc54d16d929bf6df04990
quality-controlled candidate: later official-source, hash-preserved K160
later capacity candidates: K168, K176, K180
hash layers 0-2: preserve all 256 experts
later layers 3-42: use the frozen candidate-specific expert map
production runtime: vLLM/XPU plus vllm-xpu-kernels
precision bakeoff: native MXFP4/BF16 versus symmetric group-128 INT4/BF16
initial context: 8K
workload: one active generation, never aggregate throughput
speculation: disabled until correct nonspeculative decode approaches 40-50 tok/s

The old ../deepseek-v4-flash-autoround-vllm/ packet remains rejected fit/support evidence for the oversized Intel AutoRound artifact. Do not mix its configs, model identity, results, or LocalMaxxing templates into this lane.

Current Status

Closed-lane override (2026-07-21): the best verified target-verified DSpark7 endpoint is 80.820052 tok/s; no later verified endpoint exceeded it. The lane is paused. Reopen only under the closeout’s explicit conditions.

The user explicitly authorized the frozen public K160 download while the small gates continue. The snapshot is 46 standard safetensors shards totaling 103107582016 bytes and includes MTP entirely in shard 46. Preserve the full archive, even though initial nonspeculative loading skips mtp.* tensors.

This K160 is for fast Intel bring-up, not automatic promotion. It uses one global 160-expert layout, including layers 0-2, and its unavailable calibration snapshot means its rank-discounted top-k frequency cannot be represented as true REAP. The original hash-preserving nested-pack plan remains the quality path after the official-source teacher is available.

The immediate work is deliberately small:

use the clean pinned DeepSeek-specific runtime worktrees;
prove exact-shape native MXFP4 and INT4 MoE kernels on B70;
add and test heterogeneous per-layer expert counts with dummy weights;
pass DeepSeek V4 sparse-attention/cache/mHC fixtures;
freeze ranking provenance, quality-suite identity, numerical tolerances, and storage budgets before the large download.

Promotion Rules

Every candidate must be labeled as a REAP-pruned derivative with exact K, manifest hash, packing format, source revision, and runtime identity. It must never be reported as the unmodified 284B checkpoint.

Promotion requires:

fully resident execution with no CPU/SSD expert offload;
native low-bit kernel traces and no BF16 expert expansion;
at least 3 GiB free on every GPU after warm 8K graph capture;
the fixed quality contract;
the fixed realistic cold suite with cached_tokens=0;
complete profile, repeatability, and artifact capture.

The historical rule was not to create a repro/ packet or LocalMaxxing submission until a complete full-model candidate passed those gates. The target-verified K160/DSpark7 record passed the declared cold-suite and exact gates and is now preserved in the standalone repro linked above, with the experimental-K160 provenance caveat retained.

Folder Map

benchmarks/: stage gates and later benchmark identities;
data/: fit audits, manifests, compact summaries, and hashes;
quality/: teacher/control suite and promotion contract;
results/: append-only experiment ledger;
patches/: source/config deltas, including failed attempts;
scripts/: helpers added only as their stage begins.

This site is open source. Improve this page.