RWKV-5 Eagle 7B (HF port)

head_dim stays NULL here rather than 128: config declares num_hidden_layers 32 and hidden_size 4096 but no num_attention_heads, no num_key_value_heads and no head_dim. The WKV mixer is a linear-attention recurrence with no softmax attention and no K/V projection pair, so a head dim derived as 4096/32 would be a number for a structure this model does not have. num_kv_heads, gqa_ratio and context_scaling are likewise undefined. kv_cache_bytes_per_token is 0: the recurrent state is fixed-size and context-independent. params_b 7.52 IS THE VENDOR CLAIM, not a tensor count. The HF port (RWKV/v5-Eagle-7B-HF) ships pytorch_model.bin with no safetensors index, so the count still cannot be machine-verified from tensor metadata. The native weights card (RWKV/v5-Eagle-7B-pth) and the 2024-01-29 announcement both state "Eagle 7B is a 7.52B parameter model" [[rwkv-eagle-7b-announcement]]. Dense recurrent, so active_params_b equals that total. flops_per_token_active = 2 * 7.52e9 = 1.504e10 is the schema one-pass approximation, not a profiled count; the WKV mixer is linear in context, which this scalar does not capture. Config facts are solid: model_type 'rwkv5', model_version '5_2', 32 layers, hidden_size 4096, head_size 64, context_length 4096, vocab 65536, and custom modeling code (trust_remote_code=True required). intermediate_size is null - RWKV5 has no separate MLP, which is why activation is also null. kv_cache_bytes_per_token = 0: the WKV mixer keeps a fixed-size state vector (hidden_size x head_size per layer, rescaled every rescale_every 6 layers by a decay matrix) rather than a growing KV cache, so memory is flat in context length - but unlike Mamba-2 this is an RNN-style recurrent state, so decoding is inherently sequential and prefill is not chunk-parallel in the same way. Serving caveat from the model card: this is explicitly the transformers port for experimentation and demo use, not the production path; the native RWKV libraries and RWKV-LM v5 Eagle checkpoints are the supported route, and the README notes no instruct tune exists for it yet. context_scaling is null: context length is native to the pretrained config, no rope_scaling. 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=recurrent. REQUIRED KERNEL CLASS: an RNN-style WKV mixer - a fixed-size state vector (hidden_size x head_size per layer, rescaled every rescale_every 6 layers by a decay matrix) updated sequentially, with NO chunk-parallel reformulation available. This is the worst case in the repo for the flop/ssm-sequential-scan-tensor-core-mismatch problem because unlike Mamba-2 it has no SSD matrix-multiplication reformulation, so there is nothing to recover the tensor-core utilization. ENGINE JOIN: llama.cpp lists RWKV6, RWKV6QWEN2, RWKV7 and ARWKV7 in llm_arch_is_recurrent - note the gap, this engine has NO RWKV5 entry, and the checkpoint's model_type is 'rwkv5' with model_version 5_2 [[mdl2-llamacpp-arch-class-registry]] [[pap-arch-llamacpp-arch-registry]]. So on this engine's own registry the architecture is outside the recurrent family it dispatches. vLLM's architecture table has no RWKV row of any kind [[mdl2-vllm-supported-models-2026-10]]. SGLang's generative table has no RWKV entry [[mdl2-sglang-generative-models]]. TensorRT-LLM has none [[mdl2-trtllm-supported-models]]. NO ENGINE RECORD IN THIS REPO STATES RWKV SUPPORT; the honest cell for this checkpoint is UNKNOWN in docs/13-model-compatibility.md. PRACTICAL CAVEAT ALREADY IN THIS RECORD AND REINFORCED BY THE JOIN: custom modeling code is required (trust_remote_code=True), params_b 7.52 is a vendor claim rather than safetensors metadata, and the model card states this is the transformers port for experimentation and demo use, not the production path. DEPLOYMENT (2026-10-17): the announcement points at the rwkv pip package, RWKV.cpp, and the RWKV-Runner desktop app for the native .pth, not at llama.cpp, vLLM, or ollama. No engine record in this repo states RWKV support, so those three do not run this checkpoint. There is no GGUF quant of this port to cite. CONTEXT_SCALING_METHOD (native): linear-attention recurrent (WKV) mixer: position lives in the recurrent state, not a rope ladder.