cstr/granite-speech-4.1-2b-nar-GGUF

granite-speech-4.1-2b-nar — GGUF

GGUF conversion of ibm-granite/granite-speech-4.1-2b-nar for use with CrispASR.

The NAR variant replaces the autoregressive Granite decoder with a non-autoregressive one — the LLM runs once over the full sequence with is_causal=False everywhere instead of token-by-token sampling. The slot positions in the LLM input absorb edit votes; a per-row argmax + unique_consecutive + drop-EOS gives the final transcript. Throughput is several times higher than the autoregressive variants on the same audio.

Architecturally NAR differs from base 4.1-2b in three places:

• —Encoder self-conditioning at layer 8 — the layer-8 CTC softmax is fed back into the hidden stream as a 1024-dim residual. The per-frame blank probability captured here also drives the posterior-weighted pool of the BPE auxiliary head.
• —BPE auxiliary CTC head (100353-vocab) on a posterior-pooled (window=4) view of the final hidden states. Its greedy decode initialises the LLM input text.
• —4-layer encoder hidden-state concatenation for the projector input ([layer 4, 8, 12, last] → 4×1024 = 4096-dim), vs. base's single layer and PLUS's 2-layer concat.

Files

Cosine parity (vs PyTorch BF16 reference, JFK 11 s clip)

encoder_out is the 4096-dim concatenation of layer 4, 8, 12 and the final encoder layer (the NAR architectural delta). The recommended and -f16enc files keep the encoder in F32/F16 so parity matches the F16 reference exactly. On -mini the encoder is fully Q4K — rounding error compounds across the 16-layer Conformer and the 4-way concat amplifies the worst-frame divergence to cosmin ≈ 0.10.

editing_logits raw cosine sits at ~0.956 on every Q4K variant — that's expected LLM-quantization noise across 100 K vocab logits. The argmax (`editinglogits_top1`) and the resulting transcript are unchanged: all four files reproduce the reference final text exactly.

Reference transcript: "and so, my fellow americans, ask not what your country can do for you. ask what you can do for your country."

Tested with `crispasr-diff granite-nle <model.gguf> <ref.gguf> samples/jfk.wav`

Usage with CrispASR

NAR uses a separate runtime from the autoregressive granite variants. Today it is reachable via the crispasr-diff harness and the granite_nle library directly; a granite-4.1-nar backend in the main crispasr CLI is the next step (see TODO.md).

# Bit-exact end-to-end transcribe via the diff harness
crispasr-diff granite-nle \
  granite-speech-4.1-2b-nar-q4_k.gguf \
  /path/to/ref.gguf \
  samples/jfk.wav

The library entry point is granite_nle_transcribe(ctx, samples, n_samples) in `src/granite_nle.h`; it returns a malloc'd UTF-8 string with the final transcript. There are also fine-grained accessors (compute_mel, run_encoder, run_projector, run_llm_editing) for partial-pipeline use.

Supported languages: English, French, German, Spanish, Portuguese.

Architecture

Total ~2.2 B parameters. NAR throughput is several times higher than the autoregressive variants because there is no token-by-token sampling loop.

Conversion

# Convert HF safetensors → GGUF F16
python models/convert-granite-nle-to-gguf.py \
  --input /path/to/granite-speech-4.1-2b-nar \
  --output granite-speech-4.1-2b-nar-f16.gguf

# Quantise F16 → Q4_K (encoder + projector preserved F32, LLM Q4_K)
crispasr-quantize granite-speech-4.1-2b-nar-f16.gguf \
                  granite-speech-4.1-2b-nar-q4_k.gguf q4_k

# Q4_K with F16 encoder/projector (smaller, no measurable parity loss)
CRISPASR_GRANITE_ENC_F16=1 \
crispasr-quantize granite-speech-4.1-2b-nar-f16.gguf \
                  granite-speech-4.1-2b-nar-q4_k-f16enc.gguf q4_k

# Aggressive Q4_K everywhere (encoder + projector + LLM)
CRISPASR_GRANITE_QUANT_ALL=1 \
crispasr-quantize granite-speech-4.1-2b-nar-f16.gguf \
                  granite-speech-4.1-2b-nar-q4_k-mini.gguf q4_k

The NAR converter is a separate script (convert-granite-nle-to-gguf.py) because the GGUF arch (granite_nle), tensor naming (BPE auxiliary head, 4-layer hidden capture indices) and self-conditioning metadata all differ from the autoregressive variants. The crispasr-quantize binary recognises both granite_speech and granite_nle archs and applies identical encoder/projector skip rules to both.

Licence

Apache 2.0 — same as the original ibm-granite/granite-speech-4.1-2b-nar.

Provenance and EU AI Act Art. 53 note

• —Upstream model: ibm-granite/granite-speech-4.1-2b-nar — published by ibm-granite.
• —Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• —What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• —Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• —Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.