evidence(distill): Phase 3 F-DISTILL-SMOKE-001 PASS on gx10 GB10

[noahgift]

🎉 F-DISTILL-SMOKE-001 DISCHARGED

Real distillation 1.5B Qwen2.5-Coder teacher → 0.5B Qwen2.5-Coder student on Blackwell GB10 (sm_121):

initial_loss = 7.6746
final_loss   = 7.2036   ← LESS THAN initial
62 steps, 122.7s, no errors

Phase 3 of SPEC-DISTILL-001 is COMPLETE.

What this proves

End-to-end on Blackwell with the full cascade:

• ✅ Teacher load (1.5B Qwen → 28 transformer blocks)
• ✅ Student load (0.5B Qwen → 24 transformer blocks)
• ✅ Forward pass (cuBLAS + pre-warmed PTX)
• ✅ KD loss computation (kd_step + DistillationLoss)
• ✅ Backward pass (no JIT-mid-training stream poisoning)
• ✅ Optimizer step (gradient accumulation + AdamW)
• ✅ Multi-step convergence (loss decreasing)
• ✅ Output checkpoint written (student-trained.apr/model.safetensors)

Cascade landed

#	PR	What
1	#1804 PMAT-700-B	cuBLAS prewarm skip
2	#1808 PMAT-698e	workspace cap (2048)
3	#1809 PMAT-698f	APR magic in weights loader
4	#1810 PMAT-698g	non-LoRA backward pre-warm
5	#1813 PMAT-698h	rms_norm_gamma_reduce pre-warm
6	#1815 PMAT-698i	FWD-CACHE diagnostic logging
7	#1817 PMAT-698j	THE root cause — warm! macro key
8	#1820 PMAT-698k	cache-key alignment (rope fwd, rmsnorm eps)
9	#1823 PMAT-698m	smoke setup non-degenerate batch
10	#1824	post-mortem doc
11	#1827 PMAT-698n	rmsnorm pre-warm at 1e-6 + 1e-5

Test plan

Evidence-only PR; the actual code changes already landed across the 11 PRs above. This PR captures the proof-of-success log + dispatch manifest for posterity.

🤖 Generated with Claude Code