Canary-1B-Flash (encoder checkpoint)

ENCODER-ONLY CHECKPOINT -- this repo publishes the Conformer encoder, NOT a complete ASR system. nemo_decoder_type is 'none' and nemo_model_type 'canary', with nemo_encoder_type 'nemo.collections.asr.modules.ConformerEncoder'. Canary in full pairs this encoder with a separate decoder LLM; that decoder is NOT in this checkpoint, so per-token decoder KV cannot be computed here and kv_cache_bytes_per_token is 0 for what this file actually contains. ENCODER: FastConformer (not the XL/full-attention variant), 32 layers, d_model 1024, 8 heads, encoder_ffn_dim 4096, silu, attention_dropout 0.1, encoder_layerdrop 0.1, conv_kernel_size 9, num_mel_bins 128, subsampling_conv_channels 256, subsampling_factor 8, vocab_size 1024, xscaling false, torch_dtype float32. 16 kHz input via the NeMo feature extractor. NO KV CACHE: single bidirectional encoder pass, no autoregressive decoding in this checkpoint. FLOP PROFILE: all precompute, batch-parallel, compute-bound -- the same shape as Parakeet CTC but at 32 layers instead of 42, and with 128 mel bins instead of 80. flops_per_token_active is NULL: encoder-only with no token-by-token decode. CONTRAST WITH THE OTHER NVIDIA ENCODER HERE: Parakeet CTC 1.1B is 42 layers / 80 mel bins and reaches 1.06B params; canary-1b-flash is 32 layers / 128 mel bins at 0.81B. Canary trades depth for a wider mel front-end, and it exists as a speech-TO-TEXT-TO-TEXT encoder feeding a text LLM rather than a terminal CTC model. params_b 0.811 = HF API safetensors total (811,049,984), despite the '1B' in the name. SERVICEABILITY JOIN (added 2026-10-03; the architecture x engine matrix this record was missing, with a citation per cell in docs/13-model-compatibility.md): FAMILY=speech-encoder. REQUIRED KERNEL CLASS: a Conformer/FastConformer encoder forward pass - convolution-augmented attention with relative positional encoding, which is NOT the BERT bidirectional encoder and NOT a causal decoder, and no engine record in this repo enumerates a FastConformer/NeMo ASR architecture. flop/cross-encoder-reranking and flop/embedding-batch-encoding both apply in shape (one full forward over the utterance, no cache, weights read once), but the operator mix is conv + attention, not pure GEMM. HARD JOIN RESULT: NO ENGINE RECORD IN data/engines/ STATES SUPPORT FOR THIS CLASS. The nearest evidence is indirect and negative - onnxruntime-genai's official architecture list is 'language + vision + speech' with Whisper named and NeMo absent [[onnxruntime-genai]], and the only speech-serving engines recorded here reach Whisper specifically (koboldcpp, llamafile, lemonade's whispercpp, openvino-genai, niche-h2ogpt) rather than a FastConformer ASR model. Treat engine choice for this checkpoint as UNKNOWN in docs/13-model-compatibility.md, not as unsupported: no record asserts either way. WHAT THE RECORD DOES ESTABLISH: kv_cache_bytes_per_token is 0 and that is correct - a streaming/CTC ASR encoder emits frames, it does not autoregress, so there is no growing KV cache and none of the paged-KV, prefix-cache or speculative-decode machinery applies. The serving constraint this class does impose is batching and audio-length variance, not cache. CONTEXT_SCALING_METHOD (native): FastConformer encoder with no autoregressive context ladder; no rope declaration.