Parakeet TDT 0.6B v2

NVIDIA Parakeet TDT is a CTC/transducer ASR model, not an autoregressive transformer: it has no self-attention KV cache and no causal decode, so num_kv_heads, head_dim, gqa_ratio and kv_cache_bytes_per_token are UNDEFINED for this architecture rather than unknown. Left null deliberately. VERIFICATION IS PARTIAL FOR THIS RECORD -- read before using it. ARCHITECTURE: FastConformer XL encoder (full attention, NOT the local-attention variant) with a TDT (Token-and-Duration) decoder, per the NVIDIA model card. TDT predicts both a token and its duration, so it emits multiple frames per autoregressive step -- a different latency/throughput trade from frame-synchronous CTC (parakeet-ctc-1.1b) and from token-by-token Whisper. UNVERIFIABLE FIELDS: this repo is a SINGLE parakeet-tdt-0.6b-v2.nemo archive. There is no HF-format config.json (404), the HF API reports no safetensors parameter metadata, and no per-layer dimensions are exposed in any machine-readable form. Therefore hidden_size, num_layers, num_kv_heads, head_dim, gqa_ratio, max_position_embeddings, activation, attention_variant and flops_per_token_active are ALL NULL because they COULD NOT BE VERIFIED -- not because they are absent from the model. kv_cache_bytes_per_token is NULL rather than 0: unlike the pure-encoder Parakeet CTC checkpoint, TDT does have a decoding loop, so a non-zero per-token cache exists but its size cannot be computed without the decoder depth and head geometry. Do not infer 0 from the neighbouring CTC record. ACTIVE PARAMS: the card's Model Architecture section says FastConformer-TDT with 600 million parameters and no experts, so this is a dense model and active_params_b equals params_b at 0.6. That is the card's figure, not a tensor count. params_b 0.6 IS NVIDIA'S OWN CLAIM from the model card ('This model has 600 million model parameters', and the README badge 'Params-0.6B'), NOT verified from tensor metadata -- flagged as a vendor claim per the repo contract. VERIFIED QUALITATIVE CLAIMS, all from the NVIDIA model card: English-only; punctuation, capitalization and timestamp prediction; transcription of audio segments up to 24 MINUTES IN A SINGLE PASS (vs Whisper's 30s ceiling -- this is the substantive architectural difference and it comes from FastConformer XL full attention over a long subsampled sequence); RTFx 3380 on the HF-Open-ASR leaderboard at batch size 128 (NVIDIA's number, batch 128, dataset-dependent -- not comparable to single-stream RTF). STREAMING vs OFFLINE: single-pass 24-minute capacity is an OFFLINE/long-form configuration. NVIDIA's separately released cache-aware variants (e.g. parakeet-tdt-0.6b-v3 streaming builds) are different checkpoints; this one is not a streaming model. Successor noted on the card: nvidia/parakeet-tdt-0.6b-v3, multilingual (25 European languages), verified to exist on the Hub but not recorded here. 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 stays null, not 0. The encoder itself has no growing cache, but the TDT decoder does have a decoding loop whose cache size is not published, so 0 would be a guess. None of the paged-KV, prefix-cache or speculative-decode machinery applies to the encoder pass. The serving constraint this class does impose is batching and audio-length variance, not cache. CONTEXT_SCALING_METHOD (native): FastConformer / CTC-TDT encoder: convolutional + attention encoder with no autoregressive context to extend. nvidia/parakeet-tdt-0.6b-v2 has no config.json (404).