🔬Modal Research Exchange

{
  "base_model": "Qwen/Qwen3.5-4B",
  "baseline": "dimuon-lite (DiMuon with Polar Express), beta1=0.9, grad_clip_norm=0.0. Same model, rank, data order, seed, renderer, batch size, steps and LR grid. The best-LR run of each arm is compared.",
  "batch_size": 64,
  "budget": {
    "gpu_hours": 1.8,
    "max_wall_minutes": 120
  },
  "command": "python - <<'EOF'\n# Custom part: the TriMuon config exists only on the fork; the baseline uses tinker.types.DimuonOptimizerConfig() / DimuonParams.\nimport os, tinker\nfrom tinker_cookbook.recipes.chat_sl import chat_datasets  # renderer + no_robots builder\nsc = tinker.ServiceClient(base_url=os.environ[\"TINKER_BASE_URL\"])  # fork URL for the TriMuon arm\nARM = os.environ.get(\"ARM\", \"pe\"); LR = float(os.environ.get(\"LR\", \"1e-4\"))\nopt_cfg = tinker.types.DimuonOptimizerConfig() if ARM == \"pe\" else tinker.types.TrimuonOptimizerConfig()  # Trimuon* added by the fork\ntc = sc.create_lora_training_client(base_model=\"Qwen/Qwen3.5-4B\", rank=32, optimizer=opt_cfg)\ntrain, test = chat_datasets.no_robots_batches(tc.get_tokenizer(), batch_size=64, max_length=2048, seed=0)\nfor step, batch in zip(range(80), train):\n    tc.forward_backward(batch, loss_fn=\"cross_entropy\").result()\n    P = tinker.types.DimuonParams if ARM == \"pe\" else tinker.types.TrimuonParams\n    tc.optim_step(P(learning_rate=LR, beta1=0.9, grad_clip_norm=0.0)).result()\nprint(\"eval_nll\", sum(tc.forward(b, loss_fn=\"cross_entropy\").result().metrics[\"loss:mean\"] for b in test) / len(test))\ntc.save_weights_for_sampler(name=f\"rx-trimuon-{ARM}-{LR}\")\nEOF",
  "dataset": {
    "hf_path": "HuggingFaceH4/no_robots",
    "n_examples": 5120,
    "name": "no_robots",
    "split": "train (eval on test)"
  },
  "eval": {
    "benchmark": "no_robots test split",
    "metric": "mean per-token NLL on assistant tokens",
    "n_examples": 500
  },
  "hypothesis": "With a Spindle fork that adds TriMuon, LoRA SFT on Qwen3.5-4B reaches held-out no_robots loss at least 0.01 nats lower than dimuon-lite (Polar Express) at its best LR, after the same 80 steps.",
  "lora_rank": 32,
  "max_seq_len": 2048,
  "method": "Fork modal-projects/spindle and add TriMuon as an optimizer option. It keeps DiMuon's recovered momentum and preconditioning rules. It replaces only the Polar Express polar step with two outer Polar Switch passes, then four cubic Newton-Schulz refinements folded into the rank x rank multiplier. Use the FP16 polar core and TF32 disabled, as in aaron's benchmark. Deploy the fork as a throwaway copy. Train rank-32 LoRA on Qwen/Qwen3.5-4B with a short tinker SDK SFT loop on no_robots. The loop uses the cookbook renderer for the model, forward_backward with cross_entropy, then optim_step. The treatment arm runs on the fork with TriMuon. The baseline arm runs dimuon-lite on the same fork (or shared Spindle) with identical code. Each arm sweeps LR over {5e-5, 1e-4, 2e-4}: 6 runs of 80 steps each, same data order and seed. Each run evaluates held-out loss at step 80 and logs per-step optimizer wall time. Qwen3.5-4B is a scaled-down stand-in for the source's Qwen3-8B rank-32 workload. Polar Switch and AdamW arms are follow-up specs, not part of this decision.",
  "optimizer": {
    "lr": 0.0001,
    "name": "trimuon (fork)"
  },
  "steps": 80,
  "success_criterion": "Validated if best-LR TriMuon eval NLL is at least 0.01 nats below best-LR dimuon-lite eval NLL, and TriMuon's median optimizer step time is not higher. Failed if the gap is under 0.01 nats or reversed. A rerun of the best PE config with a second seed is optional, if budget allows, to check that the 0.01 margin exceeds seed noise.",
  "title": "TriMuon vs dimuon-lite (Polar Express) for LoRA SFT on no_robots",
  "training_type": "sft"
}

[
  {
    "arm": "pe",
    "step": 80,
    "uri": "tinker://c4510b7f-55b6-4072-b10d-e3161ac8be12:train:0/weights/rx-trimuon-pe_lr0.0002"
  },
  {
    "arm": "pe",
    "step": 80,
    "uri": "tinker://55d7a67e-c109-4078-9621-321791aec6f2:train:0/weights/rx-trimuon-pe_lr0.0001"
  },
  {
    "arm": "pe",
    "step": 80,
    "uri": "tinker://e84c0575-cd19-4f8e-8a7e-6ab842e01714:train:0/weights/rx-trimuon-pe_lr5e-05"
  },
  {
    "arm": "trimuon",
    "step": 80,
    "uri": "tinker://9595714e-ffb2-46a5-a18a-728b478f7b0c:train:0/weights/rx-trimuon-trimuon_lr0.0001"
  }
]

{
  "pe": {
    "eval_nll": 2.1021892036243415,
    "median_optimizer_step_ms": 128.38237099992966
  },
  "trimuon": {
    "eval_nll": 2.1033097226039286,
    "median_optimizer_step_ms": 270.7870204999949
  }
}

base_eval_nll

2.209

effect_best_lr_nll

0.001121

• checkpoint · pe · step 80 · tinker://c4510b7f-55b6-4072-b10d-e3161ac8be12:train:0/weights/rx-trimuon-pe_lr0.0002
• checkpoint · pe · step 80 · tinker://55d7a67e-c109-4078-9621-321791aec6f2:train:0/weights/rx-trimuon-pe_lr0.0001
• checkpoint · pe · step 80 · tinker://e84c0575-cd19-4f8e-8a7e-6ab842e01714:train:0/weights/rx-trimuon-pe_lr5e-05
• checkpoint · trimuon · step 80 · tinker://9595714e-ffb2-46a5-a18a-728b478f7b0c:train:0/weights/rx-trimuon-trimuon_lr0.0001
• effective_config · effective-config.json
• code_revision · code-revision.txt

1. Nov 4, 17:42 UTC · succeeded · done: pe, trimuon · On a throwaway Spindle fork, TriMuon and dimuon-lite (Polar Express) each trained rank-32 LoRA on Qwen3.5-4B / no_robots for 80 steps at LR 5e-5, 1e-4 and 2e-4 (6 runs). At the best LR (2e-4 for both…
2. Nov 4, 17:40 UTC · running · done: pe, trimuon · All 6 runs (dimuon-lite PE vs TriMuon, LR 5e-5/1e-4/2e-4, 80 steps) finished on fork spindle-fork-run_9db47274c3; best-LR TriMuon held-out NLL 2.10331 vs PE 2.10219 (TriMuon +0.0011 worse), with TriM…
3. Nov 4, 17:18 UTC · running · Shared Spindle rejects dimuon (HTTP 400), so both arms run on fork spindle-fork-run_9db47274c3 (dimuon-lite + new TriMuon family, 1x H100 trainer). Correctness preflight passed offline, on-fork probe…
4. Nov 4, 17:10 UTC · running · Shared Spindle rejects the dimuon optimizer (HTTP 400), so both arms run on a throwaway fork (spindle-fork-run_9db47274c3) with dimuon-lite plus a new TriMuon family; offline correctness preflight pa…

adaptation_manifests

[{'arm': 'pe', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': ['80 batches of 64 no_robots prompt_ids (train.shuffle(seed=0)[0:5120]); full list in logs/*.final.json'], 'adaptation_batch_hash': '8bda2800a30081d68f62bd062260db739e51df2496456b75bc1758225960ab09'}, {'arm': 'trimuon', 'pair': 'seed0', 'warmup_batches': 0, 'adaptation_batch_ids': ['same 80 batches as pe'], 'adaptation_batch_hash': '8bda2800a30081d68f62bd062260db739e51df2496456b75bc1758225960ab09'}]

correctness_preflight

{'required': True, 'paths': [{'id': 'trimuon_optimizer_update', 'fixed_input': {'sha256': 'ffdda2c3da50a8ab7b255fd58d59f46578613c90a3750f8cd10bbc938c24c820', 'description': 'preflight/fixture.pt: torch seed 1234 LoRA factors A[32x2560], B_fc1[2048x32] (row_split 1024,1024), B_o[2560x32], with gradients and BF16 previous momentum; fresh optimizer per probe, beta1=0.9, rank 32'}, 'baseline_revision': 'modal-projects/miles 228b59d90 (dimuon-lite) dimuon.py', 'reference_revision': 'fp64 reference of the spec algorithm: unfused PS3-55 (polar-switch 6f382f1 coefficients) + 4 explicit cubic NS on the rectangular matrix, in preflight.py', 'implementation_revision': 'miles fork commit c492d1b8310f941de533bc282f8e7899f6eed531 (dimuon.py)', 'precision': 'CPU; FP32 normalization, FP16 operands with FP32 accumulation, one FP16 rounding per inner step; TF32 off during the polar core', 'normalization': 'Dimuon: lagged diagonal preconditioning, direction Frobenius-normalized to sqrt(rank), lr*(rank/32)^0.175', 'errors': {'baseline_update_A': {'max_abs_error': 0.0, 'reference_max_abs': 7.51810148358345e-06, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0}, 'baseline_update_B_fc1': {'max_abs_error': 0.0, 'reference_max_abs': 1.23586505651474e-05, 'relative_l2_error': 0.0, 'cosine_similarity': 1.0}, 'reference_update_A': {'max_abs_error': 1.0432264931872696e-08, 'reference_max_abs': 7.4132713491153066e-06, 'relative_l2_error': 0.0008565332157562118, 'cosine_similarity': 0.9999996331763589}, 'reference_update_B_fc1': {'max_abs_error': 2.2261765214593983e-08, 'reference_max_abs': 1.2032333443432177e-05, 'relative_l2_error': 0.0012576206356017146, 'cosine_similarity': 0.9999992091948885}}, 'lr_scaling': [{'learning_rate': 5e-05, 'update_norm': 0.0002828439464792609}, {'learning_rate': 0.0001, 'update_norm': 0.0005656859138980508}, {'learning_rate': 0.0002, 'update_norm': 0.0011313727591186762}]}]}

correctness_validation

{'status': 'unverified', 'required': True, 'checks': [], 'reason': 'Correctness preflight was not recorded.', 'evidence_source': 'executor report; remote execution is not attested', 'applicability': 'unknown', 'reported_verdict': 'refuted'}

evaluation

{'version': 1, 'evaluators': {'5e2ed84f0ea8e42b717cc9487d7b3fd5a399b07b37d09a430a5554c36b99eb86': {'code_sha256': '0421c9a5830b23c36f59cd50c2feadca1b9d1976b8acef6778a83af87ae0c5d4', 'prompt_template': 'tinker_cookbook renderer qwen3_5 for Qwen/Qwen3.5-4B; conversation_to_datum(max_length=2048, TrainOnWhat.ALL_ASSISTANT_MESSAGES)', 'decoding': {'temperature': 0, 'max_tokens': 0, 'stop': []}, 'parser': 'none: teacher-forced cross_entropy logprobs; loss_eval.LossEval token_mean over weight>0 tokens', 'items': {'manifest_ref': 'items.jsonl', 'manifest_sha256': 'd215ca9c57a946b482c097c771564ba02a3e117aa090ab0241c2e3bb8ca80c82'}}}, 'scores': [{'arm': 'base', 'metric': 'nll_token_mean', 'value': 2.2094380058754504, 'evaluator_id': '5e2ed84f0ea8e42b717cc9487d7b3fd5a399b07b37d09a430a5554c36b99eb86', 'seed': 0}, {'arm': 'pe', 'metric': 'nll_token_mean', 'value': 2.1021892036243415, 'evaluator_id': '5e2ed84f0ea8e42b717cc9487d7b3fd5a399b07b37d09a430a5554c36b99eb86', 'seed': 0}, {'arm': 'trimuon', 'metric': 'nll_token_mean', 'value': 2.1033097226039286, 'evaluator_id': '5e2ed84f0ea8e42b717cc9487d7b3fd5a399b07b37d09a430a5554c36b99eb86', 'seed': 0}], 'comparisons': [{'metric': 'nll_token_mean', 'arms': ['pe', 'trimuon']}]}

evaluator_validation

{'status': 'comparable', 'issues': [], 'comparisons': [{'metric': 'nll_token_mean', 'status': 'comparable', 'arms': ['base', 'pe', 'trimuon'], 'evaluator_ids': ['5e2ed84f0ea8e42b717cc9487d7b3fd5a399b07b37d09a430a5554c36b99eb86'], 'missing_arms': []}]}

feasibility

{'status': 'unknown', 'problems': ['Re-projection required: lora_rank changed or was not recorded', 'Re-projection required: dataset changed or was not recorded', 'Re-projection required: steps changed or was not recorded', 'Re-projection required: batch_size changed or was not recorded', 'Re-projection required: max_seq_len changed or was not recorded', 'Re-projection required: arms changed or was not recorded', 'Re-projection required: n_examples changed or was not recorded', 'Re-projection required: n_eval changed or was not recorded', 'Primary endpoint changed or was not recorded'], 'wall_seconds': None, 'gpu_hours': None, 'known_gpu_hours': 3.807888888888889, 'phases': {'setup': {'wall_seconds': 900.0, 'planned_units': 1, 'planned_samples': 0.0, 'gpu_hours': 0.25}, 'training': {'wall_seconds': 11840.0, 'planned_units': 480, 'planned_samples': 30720.0, 'gpu_hours': 3.2888888888888888}, 'evaluation': {'wall_seconds': 68.4, 'planned_units': 3000, 'planned_samples': 3000.0, 'gpu_hours': 0.019000000000000003}, 'recovery': {'wall_seconds': 900, 'gpu_hours': 0.25}}, 'generation_counts': None, 'generation_applicability': {'check': 'generation_audit', 'status': 'unknown', 'enforce': False, 'reasons': ['spec has no experiment_features: generated_training_data'], 'features': {'generated_training_data': None}}, 'record': {'config': {'base_model': 'Qwen/Qwen3.5-4B', 'pilot': '2 concurrent TriMuon clients x 3 steps x 64 seqs + 500-item eval'}, 'effective_config': {'base_model': 'Qwen/Qwen3.5-4B', 'arms': ['pe', 'trimuon'], 'lr_grid': [5e-05, 0.0001, 0.0002], 'steps': 80, 'batch_size': 64, 'max_seq_len': 2048, 'n_eval': 500}, 'primary_endpoint': 'mean per-token NLL on assistant tokens (no_robots test, 500)', 'source': 'pilot.json; logs/*.jsonl', 'measurements': {'setup': {'seconds': 900, 'units': 1, 'planned_units': 1, 'samples': 0, 'gpu_count': 1, 'allocation_source': 'fork config trainer_gpus_per_node=1 (H100)'}, 'training': {'seconds': 148, 'units': 6, 'planned_units': 480, 'samples': 384, 'gpu_count': 1, 'allocation_source': 'fork config'}, 'evaluation': {'seconds': 11.4, 'units': 500, 'planned_units': 3000, 'samples': 500, 'gpu_count': 1, 'allocation_source': 'fork config'}}, 'generation': None, 'recovery': {'seconds': 900, 'gpu_count': 1, 'allocation_source': 'fork config'}}, 'recorded_at': 1793814147.208957}

feasibility_required

True

initialization_provenance

{'version': 1, 'arms': {'pe': {'initialization': {'base_checkpoint': 'Qwen/Qwen3.5-4B', 'source': 'server_response', 'checkpoint_uri': 'tinker://7a2c0862c3f70072bb9645c172119bfd/weights/rx-trimuon-init-step0', 'checkpoint_immutable': True, 'checkpoint_scope': 'full_initialization', 'loaded_checkpoint_uri': 'tinker://7a2c0862c3f70072bb9645c172119bfd/weights/rx-trimuon-init-step0', 'load_succeeded': True, 'adapter_sha256': None, 'evidence_ref': 'logs/pe_*.jsonl created events', 'before_first_update': True}, 'restore_count': 1, 'optimizer_checks': [{'after_restore': 1, 'before_next_update': False, 'requested': 'dimuon', 'actual': 'dimuon', 'actual_config': {'optimizer_family/dimuon': 1.0}, 'source': 'server_response (optim_step metrics, every step)', 'evidence_ref': 'logs/pe_*.jsonl'}], 'initial_output': {'available': True, 'batch_sha256': '50b7ce3af47996f807107c50c7c0af699a17d9fcd082c38d0353d8f5eb83093e', 'output_sha256': '18240081f56bbf455d57c043b3fda21cf5f5f67ebebb8f7e3e31f75a902af34e', 'evidence_ref': 'logs/*.final.json', 'before_first_update': True}}, 'trimuon': {'initialization': {'base_checkpoint': 'Qwen/Qwen3.5-4B', 'source': 'server_response', 'checkpoint_uri': 'tinker://7a2c0862c3f70072bb9645c172119bfd/weights/rx-trimuon-init-step0', 'checkpoint_immutable': True, 'checkpoint_scope': 'full_initialization', 'loaded_checkpoint_uri': 'tinker://7a2c0862c3f70072bb9645c172119bfd/weights/rx-trimuon-init-step0', 'load_succeeded': True, 'adapter_sha256': None, 'evidence_ref': 'logs/trimuon_*.jsonl created events', 'before_first_update': True}, 'restore_count': 1, 'optimizer_checks': [{'after_restore': 1, 'before_next_update': False, 'requested': 'trimuon', 'actual': 'trimuon', 'actual_config': {'optimizer_family/trimuon': 1.0}, 'source': 'server_response (optim_step metrics, every step)', 'evidence_ref': 'logs/trimuon_*.jsonl'}], 'initial_output': {'available': True, 'batch_sha256': '50b7ce3af47996f807107c50c7c0af699a17d9fcd082c38d0353d8f5eb83093e', 'output_sha256': '18240081f56bbf455d57c043b3fda21cf5f5f67ebebb8f7e3e31f75a902af34e', 'evidence_ref': 'logs/*.final.json', 'before_first_update': True}}}, 'notes': 'Initial-output hashes match, but LoRA B=0, so they cannot show that A matches. Step-1 grad norms on the same batch were 83.452-83.463 across all 6 runs. Separately created clients differed by about 2% (2.374 vs 2.417), so this agreement is consistent with the same A plus nondeterminism from concurrent batching. The pilot restore-parity check (2 TriMuon clients from the init checkpoint, 3 identical steps) was not bitwise equal: grad norms 83.4487 vs 83.4511. The optimizer family is seen in the first optim_step response, not before it. adapter_sha256 was not measured.'}

launch

{'experiment_id': 'exp_ff293ca443', 'context_status': 'legacy_unknown', 'superseded': False, 'mode': 'ordinary', 'purpose': 'Execute the selected saved experiment version.'}

progress_updated_at

1793814099.5719247

reported_verdict

refuted

restoration_validation

{'status': 'unverified', 'required': False, 'checks': [], 'reason': 'Initialization and optimizer evidence was not required by this saved spec.', 'applicability': 'unknown'}

started_at

1793811143.2079325

verdict_report

{'effect': 0.001120518979587093, 'margin': 0.01, 'lower': 0.0008416961973258807, 'upper': 0.0014219279184608524, 'seeds': 1, 'confidence': 'weak', 'reasoning': "At the best LR, TriMuon's NLL minus PE's NLL is +0.00112 nats (paired item-bootstrap 95% CI [0.00084, 0.00142]). The success criterion needed -0.01 or lower, so the gap is in the wrong direction. TriMuon's median server-side optimizer step was also 271 ms against 128 ms for PE, so the timing condition fails too. Confidence is weak for three reasons: there is one training seed (the CI covers test items only); both arms picked the top edge of the LR grid (2e-4), so neither arm is tuned; and TriMuon is my own reconstruction (the upstream miles PR #27 code was withdrawn), run eagerly in PyTorch without fused kernels.", 'grid_edge': True}