## Problem
The two `moe_fp32` definitions (`moe_fp32_e60_k4_d2048_ff1408`, `moe_fp32_e64_k8_d2048_ff1024`) were **orphaned** after the bf16 migration: the definitions exist but had **no solutions** (no `reference-scalar`, no `baseline-llamacpp-arm`), so they evaluate to `NO_REFERENCE` and cannot run.
## Fix
Restore both solutions by adapting the matching `moe_bf16` ones (identical SwiGLU-MoE logic, only the input dtype differs):
- **reference-scalar**: `uint16_t`(bf16)-> `float` inputs; drop `bf16_to_f32` widening (read fp32 directly).
- **ggml baseline**: `GGML_TYPE_BF16` -> `GGML_TYPE_F32`.
## Verification
- reference-scalar: compiled (clang -O2 -std=c++14), run via ctypes, output **matches `Definition.reference`** (max_rel_err = 1.9e-6).
- ggml baseline: **compiles against ggml** on Graviton4 (clang-18, ggml headers) for both defs.
## Note
Restores fp32 MoE parallel to `gemm_fp32`. If the bf16 migration intended to **deprecate** fp32 MoE instead, the alternative is to delete the …
Rewrite llama.cpp q8_0 reference-scalar solution signature, now compatible with bench/dataset/raw.py
ArtificialRay
6b3c7a1
Add kernel in different precision:
ArtificialRay
c09d00d
- CV models: FP32 kernels and int8 (w8a8ch) kernels
- LLM models: BF16 kernels and int4 (Q4_K) kernels
Full precision kernels (fp32 and bf16) are ready to benchmark while low-bit kernels have correctness issue that may need a new evaluator
