code-HFD / paper_context_v4 /architecture_spec.json
Tinghong-Ye22's picture
Add files using upload-large-folder tool
5d9f970 verified
Raw History Blame Contribute Delete
3.72 kB
{
"format": "hfd_paper_architecture_v4",
"date": "2026-10-08",
"code_snapshot": "9fbfa32",
"recipe": "a1_c_deferred",
"action_representation": "root40_v1",
"policy_backbone": {
"name": "One branched Qwen-derived interaction backbone",
"indexing": "zero_based",
"path_depth": 36,
"hidden": 2048,
"q_heads": 16,
"kv_heads": 2,
"head_dim": 128,
"ffn": 11008,
"shared_layers": [
0,
27
],
"body_private_layers": [
28,
35
],
"qa_private_layers": [
28,
35
],
"unique_decoder_layer_parameter_sets": 44,
"shared_base_embedding": true,
"shared_prefix_frozen": true,
"compute_reuse": "only if current inputs, positions, masks, and prefix histories are equivalent",
"mutable_kv_containers": "private body and QA",
"body_policy": "tail8 and final norm trainable",
"qa_policy": "frozen tail8 base with rank16 alpha16 LoRA on q/k/v/o and gate/up/down"
},
"cross": {
"layer_index": 34,
"position": "after native self-attention residual, before native MLP",
"type": "two directed independent cross-attention modules",
"donor": "current packet pre-cross residual, detached",
"inner_dim": 256,
"heads": 8,
"head_dim": 32,
"max_memory_tokens": 32,
"mask": "donor valid and original donor index <= original query index",
"gate": "tanh(scalar), initial zero",
"parameters_both": 4194306
},
"control": {
"label_order": [
"listen",
"motion",
"idle"
],
"tokens": [
"<|listen|>",
"<|motion|>",
"<|idle|>"
],
"verified_run_token_ids": [
151665,
151666,
151667
],
"rows_shape": [
3,
2048
],
"prediction": "last_valid_body_hidden @ rows.T",
"action_condition": "detach(rows[argmax])",
"write": "predicted token prepended to NEXT AV packet in BOTH histories",
"source_reset_clears_pending": true,
"anchor_weight": 0.2
},
"action": {
"action_dim": 40,
"state_dim": 40,
"absolute_pose_dim": 36,
"width": 1536,
"blocks": 6,
"heads": 24,
"head_dim": 64,
"ffn": 6144,
"context_order": [
"body_current_tokens",
"detached_control_row",
"observed_state"
],
"context_length": "L+2",
"noisy_action_shape": [
"B",
30,
40
],
"input_mlp": [
40,
40,
1536
],
"input_precision": "FP32 input MLP and position addition; rest mixed precision",
"output": "flow velocity Bx30x40",
"solver_steps": 10,
"predict_frames": 30,
"execute_frames": 20,
"rtc_frames": 10,
"rtc_rebase": true,
"root40_slices": {
"joint_abs": [
0,
29
],
"delta_anchor_xy": [
29,
31
],
"rotation_anchor_6d": [
31,
37
],
"world_z": [
37,
38
],
"contacts": [
38,
40
]
},
"parameter_count": 500108048
},
"clock": {
"sample_rate": 16000,
"audio_window_seconds": 0.6,
"semantic_stride_seconds": 0.4,
"camera_hz": 8,
"action_hz": 50,
"training_action_stride_frames": 1,
"inference_action_stride_frames": 20
},
"speech": {
"policy_answer": "QA private snapshot autoregressive text",
"auxiliary_renderer": "separate frozen Voice Thinker (text only), Talker, Token2Wav",
"renderer_reads_av": false,
"codec_loss_trained": false,
"pcm_hz": 24000
},
"measured_policy_parameters": {
"source_run": "hfd.gpus2.2610072107",
"unique": 6143554244,
"trainable": 1121244740,
"excludes": "separately instantiated inference speech renderer"
}
}