keryx-miner-supr 0.15.4 — H15 private inference (traced answers), tier 4 fix

Update from 0.15.3 if a card of yours serves tier 4 (Kimi-Linear-48B), and on Windows. 0.15.4 is 0.15.3 plus three fixes found while testing 0.15.3 on our own rigs:

• Tier 4 can be served. 0.15.3's H15 self-test refused tier 4 (Kimi-Linear-48B), because its MLA cache fell back to an f16/f16 cache that the trace cannot commit. Those cards correctly did not mine tier 4. 0.15.4 uses a Q8_0 K / f16 V cache, like keryx v0.5.8.
• Declare only after the traced self-test. A card now tells the pool it serves a tier only after its traced self-test has passed. Before, the plain warm-up answer counted, about two seconds earlier.
• Windows. The Windows NVIDIA build is fixed; 0.15.3 shipped no Windows package. The Windows NVIDIA and AMD zips are still building and will be attached to this release shortly.

Everything below applies to 0.15.4 as it did to 0.15.3.

H15 private inference (since 0.15.3)

Update now if you mine solo or serve inference. Since the H15 gate (mainnet DAA 140,115,000, 2026-10-30) keryxd 1.6.6 requires every private inference answer to carry a verifiable trace, and requesters seal their requests with key schedule 2. 0.15.0–0.15.2 mine H15 blocks correctly, but they could not open schedule-2 requests and their answers carried no trace, which the node rejects — and a missed answer costs a service-bond strike. 0.15.3 serves H15 private requests the way keryxd 1.6.6 expects. Mining itself is unchanged.

• Key schedule 2. Private requests are opened with key schedule 2, falling back to schedule 1. The answer is sealed with the per-responder key of whichever schedule opened it.
• Traced answers. Every private answer is generated with the canonical inference trace armed: the activations of every layer, the KV cache, the attention softmax mass and the recurrent state are committed, together with the token ids of the answer. The trace commitment is attached to the AiResponse before it is signed. The response is only sent once its trace is safely on disk.
• Audits and openings. The miner follows the node's draw for every answer it sent: the seed block, whether it is audited, and the deadline. With keryxd 1.6.6 or newer it also uses the node's own list of awaited openings. It then replays the drawn positions on the card that generated the answer, checks every opening with the node's own verifier code, and submits it. Pending traces survive restarts and reconnects (<escrow state>.traces/), and the chain is rebuilt after a restart, so no audit is lost.
• Self-test before mining. Before a card mines its tier, it must answer a private request the H15 way. It runs a traced generation, then checks the openings (trace, attention, recurrent and residual pieces) with the node's verifiers for two seeds. This takes a few seconds. A card or build that cannot do this does not mine, and the log names the card and the reason: mining a tier you cannot serve only collects strikes.
• Engine. libkeryx-llama.so / keryx-llama.dll is now engine ABI 4. It adds the traced-generation and sequence-leaf functions and the verifier-exact attention and RoPE kernels of keryx v0.5.8. Always use the engine that comes with this release, next to the miner binary.
• Pool mining is unchanged for pools that do not ask for traces. A pool can now ask for a traced answer per request; the rig then returns the trace commitment with the answer and keeps the trace for the pool's audits (9 h, on disk). mining.hello / mining.telemetry report "trace": 1 once the self-test passed.

Platform notes

• NVIDIA (Linux modern/legacy, Windows, Docker): full H15 support, including traced answers and openings.
• AMD (OpenCL/Vulkan) and Apple Silicon (Metal): these backends have no trace engine, so they cannot answer H15 private requests. After the gate they do not mine; they log why instead.
• Linux packages run on Ubuntu 20.04 and newer (glibc 2.31+), including current HiveOS, mmpos and SMOS images.

Validation

Tested on our rigs with the published 0.15.3 packages and the 0.15.4 builds: RTX 5090 (tiers 2 and 3), RTX 5080 (tiers 0, 1 and 2), RTX 3070 (tier 0, modern and legacy builds) and CMP 170HX (tiers 1, 3 and 4; tier 4 with the 0.15.4 build). On testnet, both key schedules were answered with traces and accepted, and their audit openings were accepted on-chain. On mainnet, through the pool, a traced answer was accepted. Same-model replays reproduce across hosts and driver versions (RTX 3070 ↔ RTX 3070, RTX 5090 ↔ RTX 5090). AMD and Apple builds refuse to mine, as described above.

Validated end to end against keryxd 1.6.6 on testnet. Schedule-2 and schedule-1 private requests were opened and answered with traces, and the node accepted the responses. Their audits were opened (trace, attention, recurrent and residual pieces, plus a later attention phase) and every opening was accepted on-chain. The requester decrypted both answers. An audit that came in after a miner restart was opened from the on-disk store. Unit tests include the node's own verifier and private-inference vectors.

Downloads

Use the matching SHA256SUMS file to verify downloads.