Strix Halo TM64 DeepSeek V4 Vulkan b54f16e1-3cacee17

IMPORTANT: FP8 KV warning

Do not use FP8 KV cache when planning to exceed 100K context. Use F16 KV for long-context operation.

Validated Strix Halo Vulkan release for the custom DeepSeek V4 TM64 target and selective-Q4 DSpark drafter.

Model

• Target GGUF: erayyapagci/DeepSeek-V4-Flash-IQ2XXS-TM64-Q4KAttention-v1
• Models are not bundled in this release.

Build provenance

• Source commit prepared for this fork: 4fc4f230f6cc7735dc05cd403a8eb73ccd64813c
• Local source base: b54f16e18ba3bd6e34b87a1ef827ca985307cb5b
• Accepted patch SHA-256: 3cacee17358f178f4c2a6c02c95c61426fec52e9f04db2b05a2d1de28c4cd623
• Runtime archive SHA-256: 2bd5eb87c7a855959bbbc6d652378a025a5e6c9bd38a6b9837ceaf963cae46da
• Source archive SHA-256: 28e0964545eef6b515b1ba60cc371356245df17e779da3cc83d2994a58988bf1
• Build: Release, Vulkan/RADV, Linux x86-64, AMD Strix Halo/gfx1151

The complete patched source tree is provided as a release asset. The runtime archive also includes the accepted patch, exact diff, CMake configuration, provenance, checksums, and validation logs.

Included

• llama-server, llama-cli, llama-bench, quantizer, tests, and common tools
• Bundled validated RADV runtime
• Tuned bin/run-server.sh reproducing the production configuration
• NP1/NP2/NP3 support with state isolation
• F16 KV and F16 lightning-indexer defaults
• FP8-over-100K launcher safety gate
• Optional slot persistence and systemd example
• Complete patched source snapshot

See the packaged README.md for tuned options, setup, safety constraints, and rollback controls.

Validated no-speculation benchmark snapshot

Metric	0K	64K
2,048-token prefill	260.90 tok/s	199.04 tok/s
128-token generation	20.62 tok/s	16.82 tok/s