KV-FP8 W1's read-side comments anchor `scaled_vec_conversion<float, uint8_t>` at `quant_utils.cuh:302-308`, which at the pin is the identity primary template plus the fp8->HALF specialization; the real one is `:419-429`

[localai-bot]

Row: KV-FP8

Three KV-FP8 W1 read-side comments cite scaled_vec_conversion<float, uint8_t> at csrc/quantization/w8a8/fp8/nvidia/quant_utils.cuh:302-308. At parity pin 5559679229bc961848b121ccdeaa8fa5d79bec98 that range holds no such thing. Lines 301-305 are the GENERIC primary template (template <typename Tout, typename Tin> ... { return x; }) and 307-314 are the // fp8 -> half <uint16_t, uint8_t> specialization, so 302-308 straddles a template that returns its input unchanged and a conversion into the WRONG output type. The <float, uint8_t> specialization the three comments describe — __nv_cvt_fp8_to_halfraw then half_to_float(tmp) * scale — is at :419-429, under the // fp8 -> float label at :418.

Sites, all landed by W1:

• include/vt/fp8_kv.h:92 — above LoadKvFp8E4M3
• include/vt/ops.h:1129 — beside PagedAttentionArgs::k_scale/v_scale
• src/vt/cpu/cpu_paged_attn.cpp:164 — the CPU read dequant

W2 (PR #1606, issue #1593) carried four copies of the same wrong anchor into src/vt/cuda/cuda_paged_attn.cu, tests/vt/test_cuda_fp8_kv_cache.cpp and .agents/specs/fp8-kv-cache.md. Those four are CORRECTED in that pull request. The three above are outside its authority, so they are filed rather than fixed in flow.

Same file, second anchor, same change: Fp8KVCacheDataType is cited at csrc/attention/dtype_fp8.cuh:9-13, which is the #include <cuda_fp8.h> guard. The enum is at :15-19. Sites: include/vt/fp8_kv.h:5 and :30.

Third item, a different kind. .agents/engine-matrix.md (KV-FP8 row) and .agents/quantization-matrix.md (QUANT-KV-FP8 row) both say the W2 CUDA translation units are UNCOMPILED. CI job cuda-fat-build built them for 80;86;87;89;90a;100a;103a;110;120a;121a under -Werror=all-warnings and passed on 4d71e776efc18cb5e61a26e642ddad8de5339134 (run 32495320287, job 96812232428). What remains true is that nothing has been EXECUTED on a device, because that job configures -DVLLM_CPP_BUILD_TESTS=OFF. Both clauses need the narrower statement.

Why this matters beyond tidiness: an upstream anchor is how the next reader checks a port against the oracle. One that lands on a return x; primary template invites the conclusion that the port is unfaithful, or worse, that the identity template IS the reference. .agents/specs/fp8-kv-cache.md ## Owed names this issue.