{ "source_commits": { "llm-compressor": "c892a93dad085a3a6aee7f7c72a8ac301c1e5ee9", "compressed-tensors": "e0b43ebab47e4ef0ecb530a499589ba9b55de16f" }, "runtime": { "torch": { "distribution": "2.13.0+cu130", "runtime": "2.13.0+cu130", "path": "/usr/local/lib/python3.12/dist-packages/torch/__init__.py" }, "transformers": { "distribution": "5.17.0", "runtime": "5.17.0", "path": "/usr/local/lib/python3.12/dist-packages/transformers/__init__.py" }, "vllm": { "distribution": "0.30.0", "runtime": "0.26.0", "path": "/usr/local/lib/python3.12/dist-packages/vllm/__init__.py" }, "accelerate": { "distribution": "1.15.0", "runtime": "1.15.0", "path": "/usr/local/lib/python3.12/dist-packages/accelerate/__init__.py" } }, "runtime_note": "vLLM distribution metadata is 0.30.0 but runtime version is 0.26.0; results are specific to installed WDC build.", "training": { "eval_loss": 0.02748331055045128, "eval_runtime": 0.0257, "eval_samples_per_second": 311.39, "eval_steps_per_second": 77.847, "epoch": 100.0, "perplexity": 1.0278644604720464, "global_step": 200, "dataset": "create-tiny-model COPYPASTAS", "mtp_trained": false }, "limitations": [ "MTP layer untrained", "Toy perplexity evaluated on training texts", "Text-only, single GPU, TP=1, eager, one speculative token", "Does not resolve sparse Inkling MoE export issue", "Fully quantized MTP E2E remains failing" ], "fixture": { "base_model": "thinkingmachines/Inkling", "base_revision": "828496eeae4c243ff1a22f7f28ff83694f2f7bc9", "tokenizer_model": "inference-optimization/Inkling-0.92B-MTP", "tokenizer_revision": "22b31818016a1560cfeb47caaa3296cafa2f43df", "seed": 0, "parameters": 54136067, "weight_bytes": 108279142, "mtp_tensors": 20, "all_keys_in_original_checkpoint": true, "mtp_keys": [ "model.mtp.layers.0.embed_norm.weight", "model.mtp.layers.0.hidden_norm.weight", "model.mtp.layers.0.input_proj.weight", "model.mtp.layers.0.transformer_block.attn.k_norm.weight", "model.mtp.layers.0.transformer_block.attn.k_sconv.weight", "model.mtp.layers.0.transformer_block.attn.q_norm.weight", "model.mtp.layers.0.transformer_block.attn.rel_logits_proj.proj", "model.mtp.layers.0.transformer_block.attn.v_sconv.weight", "model.mtp.layers.0.transformer_block.attn.wk_dv.weight", "model.mtp.layers.0.transformer_block.attn.wo_ud.weight", "model.mtp.layers.0.transformer_block.attn.wq_du.weight", "model.mtp.layers.0.transformer_block.attn.wr_du.weight", "model.mtp.layers.0.transformer_block.attn.wv_dv.weight", "model.mtp.layers.0.transformer_block.attn_norm.weight", "model.mtp.layers.0.transformer_block.attn_sconv.weight", "model.mtp.layers.0.transformer_block.mlp.global_scale", "model.mtp.layers.0.transformer_block.mlp.w13_dn.weight", "model.mtp.layers.0.transformer_block.mlp.w2_md.weight", "model.mtp.layers.0.transformer_block.mlp_norm.weight", "model.mtp.layers.0.transformer_block.mlp_sconv.weight" ], "backbone": "toy-trained using create-tiny-model skill", "mtp": "deterministically initialized, not trained" }, "tested_weight_sha256": "3f37052495cee65d0c424f9dad0679b4e32f2daca26d8ae7431d3db827029ac1", "purpose": "Test detection, loading, quantization, and export of explicit model.mtp checkpoint layers. Separate BF16 MTP serving controls pass; fully quantized MTP serving is blocked by the tested vLLM loader.", "validation": { "bf16_mtp_serving": "passed", "fp8_backbone_bf16_mtp_serving": "passed", "fp8_mtp_export": "passed", "fp8_mtp_serving": "failed: vLLM Inkling MTP quant_config=None", "copied_mtp_serving": "passed: oneshot saved FP8 backbone, all 20 MTP tensors exactly equal to source, both prompts generated tokens, positive MTP draft counter", "toy_prompt": "The capital of France is -> Paris. Paris is the largest city in France..." } }