Parakeet CTC 1.1B

ENCODER-ONLY. THIS MODEL HAS NO DECODER AND THEREFORE NO KV CACHE WHATSOEVER -- kv_cache_bytes_per_token is 0 BY CONSTRUCTION, not because the value is unknown. This is the cleanest possible contrast with Whisper and it is why CTC models are the cheap option for offline ASR. ENCODER (is the whole model): FastConformer, 42 layers, hidden_size 1024, 8 heads, intermediate 4096, silu, attention_bias true, conv_kernel_size 9 (depthwise ConvModule in each layer instead of an MLP), subsampling_conv_channels 256, subsampling_conv_stride 2, subsampling_factor 8, num_mel_bins 80, max_position_embeddings 5000, ctc_zero_infinity true. Preprocessor: ParakeetFeatureExtractor, 16 kHz, n_fft 512, hop 160, win 400, feature_size 80, preemphasis 0.97. NO ATTENTION CACHE: the encoder runs once over the subsampled frame sequence and CTC emits labels by a frame-synchronous greedy/beam argmax over the 1,025-token vocabulary (vocab_size 1025). There is no autoregressive step, so there is nothing to cache and no per-token KV at all -- a 1-hour recording costs the same resident memory as a 1-second one. Contrast the encoder-decoder case where every in-flight request pins hundreds of MiB of cross-attention KV. FLOP PROFILE: entirely precompute. Same qualitative shape as Whisper's encoder (bidirectional, parallel, compute-bound, batch-friendly) with NO decode phase at all. flops_per_token_active is NULL rather than 2*params: the output is frame-synchronous CTC, not one token per step, so a per-token FLOP figure has no meaning here. FastConformer normally uses LOCAL (sliding-window) attention for its encoder, which is what makes it fast on long audio; this HF-format config does not expose the local-attention window size, so attention_variant records 'MHA' as the only verifiable fact and the local-attention window is left unrecorded rather than guessed. params_b 1.063 = HF API safetensors total (1,062,626,305). STREAMING: not supported by this architecture (bidirectional encoder), but unlike Whisper there is no fixed 30s window either -- max_position_embeddings 5000 subsampled frames implies a much longer single-pass capacity than Whisper's 1,500. 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; no autoregressive context ladder, no rope declaration.