| model: | |
| repository: "ByteDance/Ouro-1.4B" | |
| revision: "e3b1e0993b1231a51d6069a870476dda4162c00a" | |
| config_sha256: "3ad25a6a186865de4cbda24d37c514e8488d8fdf4536fa47c10ec0efd5c8970a" | |
| modeling_ouro_sha256: "59e7523e9fe23f75bde78bb914d1a4d24f3997ae02bca510bbaef947356c447b" | |
| physical_layers: 24 | |
| recurrent_loops: 4 | |
| hidden_size: 2048 | |
| intermediate_size: 5632 | |
| dtype: bfloat16 | |
| groups: | |
| attention_qkv: | |
| hf_weights: [self_attn.q_proj, self_attn.k_proj, self_attn.v_proj] | |
| vllm_consumer: self_attn.qkv_proj | |
| transform_site: self_attn.ln_trans | |
| attention_output: | |
| hf_weights: [self_attn.o_proj] | |
| vllm_consumer: self_attn.o_proj | |
| transform_site: self_attn.o_trans | |
| mlp_up_gate: | |
| hf_weights: [mlp.gate_proj, mlp.up_proj] | |
| vllm_consumer: mlp.gate_up_proj | |
| transform_site: mlp.up_gate_trans | |
| mlp_down: | |
| hf_weights: [mlp.down_proj] | |
| vllm_consumer: mlp.down_proj | |
| transform_site: mlp.down_trans | |
| runtime_status: | |
| adapter_contract: implemented | |
| vllm_hook_integration: not_implemented | |
| smoke_generation: not_run | |
| required_vllm_integration_points: | |
| artifact_load: "OuroForVLLM.__init__ must load and validate the isolated LoopQ artifact." | |
| qkv_input: "OuroAttention.forward immediately before self.qkv_proj." | |
| attention_output_input: "OuroAttention.forward immediately before self.o_proj." | |
| mlp_up_gate_input: "OuroMLP.forward immediately before self.gate_up_proj." | |
| mlp_down_input: "OuroMLP.forward immediately before self.down_proj." | |
| weight_load: >- | |
| OuroForVLLM.load_weights must install the offline folded/RTN W4 projection | |
| representation without invoking the NVFP4 policy. | |
| selected_weight_dispatch: >- | |
| Implemented for eager validation: retain one shared base and four packed | |
| variants only for each of the four SLT-selected groups, dispatching by | |
| recurrence index 0..3. | |
| transition: >- | |
| OuroModel.forward must apply CTA after completed loops 0, 1, and 2 only, | |
| after the recurrent boundary norm and before the next loop consumes state. | |
| smoke_gate: >- | |
| Load a calibrated artifact and complete vLLM generation with all foreign | |
| quantization environment policies disabled. | |
| smoke_command: >- | |
| CUDA_VISIBLE_DEVICES=0 PYTHONPATH=./loopQ | |
| python loopQ/scripts/smoke_ouro_vllm.py | |
| --artifact models/loopq/ouro_w4a4.pt | |
| --gpu 0 | |