A single-MI300X DeepSeek V4 Flash deployment with two named correctness fixes

Sources: primary, discussion
Summary: The repository documents a single-MI300X deployment of DeepSeek V4 Flash, pins every runtime artifact by SHA-256, and records upstream base revisions for each diff. The first of two named fixes is correctness rather than throughput: an MXFP4 mixture-of-experts kernel masked its bitmatrix padding lanes against the global tensor bound instead of the logical block size, which corrupted routing under load and produced near-match tool names and forgotten schemas on long prompts. The second is that MI300X implements the AMD FNUZ variant of E4M3 while MI325X and newer use OCP-standard FP8, so a kernel that assumes OCP semantics can be wrong by a factor of two in the scale domain.
Why it matters: Both traps degrade output quality instead of failing loudly, which is the failure mode anyone porting an FP8 model to CDNA3 is least likely to catch.