[lkarlslund]

Adds an experimental edda_asr community family for Edda v0.2 large, a Danish Whisper-large-v3 fine-tune, and for edda-v0.2-large-hurtig, which adds syv.ai's 170M-parameter DFlash2 drafter for lossless speculative decoding.

What it does

• Speculative decoding. The drafter proposes up to 15 tokens per step from edda decoder layers 1/8/15/22/29, its own cross-attention and DFlash2's grouped dynamic causal convolutions and candidate selector; edda verifies the block in one pass. 5.24 tokens per edda pass on FLEURS da_dk (the model card reports 5.56).
• Greedy decoding for the plain edda release, or with speculative=false.
• Beam search (num_beams, length_penalty) following Transformers' _beam_search, using a token-tree KV cache so pruned beams need no cache copies.
• Silence-based chunking for long audio, pull-streaming output per chunk, server support, GGUF conversion with the drafter under a drafter. prefix.
• Docs in docs/models/edda_asr.md, model spec, model-manager packages, CLI path cases, WebUI catalog entry.

The encoder is model-local rather than the shared Whisper embedding: it packs QKV and uses flash attention over F16 K/V, which took it from ~430 ms to ~285 ms on gfx1151.

Results (AMD Radeon 8060S, HIP, 11.3 s mean FLEURS clip)

Mode	Decoder	Total
Greedy, F16	597 ms	862 ms
Speculative, F16	193 ms	519 ms
Speculative, Q8_0 decoder + F16 encoder	152 ms	428 ms
Beam 5, F16	929 ms	1234 ms

Against PyTorch FP32 references: FLEURS 60 clips 59/60 identical for both speculative and greedy (the misses are near-ties on a name spelling, as the model card reports for fp16); beam search 10/10 on the clips where beam differs from greedy; NST 5/5 for every GGUF variant, beam search and the CPU backend; 180 s long-form at 0.0% WER. Full table in the docs.

Notes for review

• It is a separate family rather than an extension of crisperwhisper, because the drafter needs its own decoder loop and edda needs its own prompt tokens. Happy to restructure if you prefer otherwise.
• No audio.cpp GGUF is published yet, so the default package is marked "convert locally".
• webui/native/dist is not rebuilt: the committed bundle is already behind its sources, and a rebuild would pull unrelated changes into this PR.
• Not tested on CUDA (the test card was fully occupied).
• model_spec_system_test and tools/check_loader_catalog_sync.py pass.

[0xShug0]

@lkarlslund Thanks for contributing a new model! One issue: with edda_asr.drafter=false, the code still calls std::clamp with an invalid range: minimum 2, maximum 1. The normal release build works, but the code can crash when C++ standard-library assertions are enabled. Please skip the clamp calculation when the drafter is disabled and use block size 1 directly.

Update: found another issue --- corrupted drafter weights are silently ignored. A damaged drafter.safetensors still succeeds with drafter=true, reporting drafter=absent and switching to greedy. It should throw an exception to make the error explicit.

[lkarlslund]

Thanks for catching these, both fixed in 4efd331:

• With the drafter disabled the draft block is now 1 directly, so the clamp to [2, 1] never runs.
• A drafter.safetensors that can't be parsed or lacks fc.weight now throws ("Edda could not read the drafter weights: …") instead of reporting drafter=absent. A missing drafter file, or a GGUF converted without the drafter/ namespace, still decodes greedily, since that's the plain edda release. To tell the two apart I added a small ResourceBundle::has_tensor_source().

Tested with a valid, garbage, truncated and missing drafter, a GGUF with and without the drafter, and drafter=false.