fix: correct 9 critical/high bugs found via multi-agent audit

[imitater-dou]

Summary

This PR fixes 9 confirmed bugs discovered through a multi-agent static analysis of the codebase, all verified by direct source inspection before patching.

🔴 Runtime Crashes (will crash on first use)

• specforge/core/loss.py — LogSoftmaxLoss.backward returned 5 values for 3 inputs (forward takes logits, target, position_mask). PyTorch autograd raises RuntimeError on every backward pass.

• specforge/distributed.py — Gather.backward returned 6 values for 5 inputs (one extra trailing None). Crashes any distributed training run that uses sequence-parallel gather.

🔴 Silent Training Correctness Bugs

• specforge/data/preprocessing.py — USP position_ids were identical for all Ulysses ranks within the same ring group (ring_start = ring_rank * ring_chunk ignores ulysses_rank). With sp_ulysses_size > 1, all Ulysses peers claimed the same global position ID range, producing wrong attention patterns. Fixed by offsetting by ulysses_rank * usp_chunk_size.

• scripts/train_eagle3.py — draft_model.eval() was set at the start of the evaluation loop but never reversed. All training steps after the first eval checkpoint ran with dropout disabled. Also, save_interval was not scaled by draft_accumulation_steps (unlike log_interval and eval_interval), causing 4× more frequent saves than intended.

• scripts/train_dflash.py — On resume, only the LR scheduler state was restored (optimizer.scheduler.load_state_dict); the AdamW optimizer's first/second moment estimates (exp_avg, exp_avg_sq) were silently discarded, producing a loss spike at every resume. Fixed by calling optimizer.load_state_dict(resume_state) which restores both.

• specforge/lr_scheduler.py — WarmupDelayerScheduler.step() passed epoch - warmup_epochs to the after-scheduler, missing the delay_epochs subtraction. The cosine schedule started at a non-zero offset, causing a discontinuous LR jump after the warmup+delay phase.

🔴 Logic Bugs in Data Pipeline

• specforge/data/parse.py — HarmonyParser used or instead of and in the first-message guard, making the condition always True. Every conversation unconditionally received the reasoning effort prefix, including those that already had a system message.

• specforge/data/template.py — TemplateRegistry.register assert was logically inverted: assert (not override and ...) raised AssertionError whenever override=True was passed, making the override parameter non-functional.

🟠 High Severity

• specforge/core/eagle3.py — _compute_metric_acc used loss_mask.sum() as the denominator while the numerator was masked by position_mask (which excludes out-of-vocabulary positions). This produced systematically deflated accuracy metrics throughout training. Fixed by using position_mask.sum() as denominator.

• specforge/modeling/target/eagle3_target_model.py — padding(target_out, left=False) was called unconditionally even when target_out is None (the case when return_logits=False). This raised AttributeError on None.__getitem__. Fixed with a None guard.

Test plan

• Verify LogSoftmaxLoss backward runs without RuntimeError on a minimal synthetic batch
• Verify Gather.backward runs without RuntimeError in a 2-GPU USP setup
• Verify process_data_usp produces distinct position_ids for each SP rank when sp_ulysses_size > 1
• Verify training resumes correctly (loss continues from pre-resume level, no spike)
• Verify draft_model.train() mode is active during training steps after an eval checkpoint
• Verify TemplateRegistry.register(name, tpl, override=True) no longer raises
• Verify HarmonyParser does not prepend reasoning effort prefix when the first message is a system message

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!