[ybenyedder]

Summary

• EAGLE3 standalone draft support (--mtp-model): loads EAGLE3 GGUF or auto-converts .safetensors + config.json (including angelslim, tengyunw, and RedHatAI/speculators transformer_layer_config / layers.0.* aliases) in pure Rust (src/gguf.rs, src/model.rs), with Vulkan graph capture, d2t remap, and non-throttled adaptive MTP depth (src/gpu_model.rs).
• Hybrid N-gram + EAGLE3 speculative cascade (--spec-type): ngram-simple, ngram-map-k, ngram-map-k4v, ngram-mod, ngram-cache (src/ngram.rs) with --lookup-cache-static / --lookup-cache-dynamic and exact-6-gram priority.
• RDNA3.5 (gfx1150) / Vulkan optimizations: disable harmful MoE L2 prefetch on gfx11xx/iGPU, honor STRATA_EXPERT_MB with multi-pool fallback, rcp_d wave32 rounding fix (common.glsl), cooperative Q3_K/Q6_K/IQ3_XXS decode shaders, single-submit command-buffer replay, and rel_l2 <= 1e-6 tolerance on sliced attn_prefill QSA mask test for AMD v_wmma 1-ULP rounding.

Verification

• cargo test --release --lib --bins --tests: 238 passed, 0 failed, 5 ignored on AMD Radeon 890M (RADV STRIX1).

[Vasyl-Trefilov]

@ybenyedder TYSM for contribute! can you please test it on main to make sure that it's working.

[ybenyedder]

@Vasyl-Trefilov Tested b99db7d on my Radeon 890M (gfx1150). 18/19 test suites pass, only tests/kernels.rs (vulkan_type_matrix and gemv_partial_superblocks_and_q2_0) is still failing because quant_xq runs in wave32 on AMD too, so rcp_d ends up hitting fdiv instead of double division.

I put up a quick PR (#37) to enable double precision rcp_d on AMD wave32 without touching the NVIDIA path. With that, all 19/19 pass on my end.