fix(distill): warm! macro used hardcoded "silu_forward" key — THE Blackwell root cause (PMAT-698j)#1817
Merged
Conversation
…" key (PMAT-698j)
THE root-cause bug behind the entire Phase 3 cuda dispatch cascade
(PMAT-698e..i, 6 prior PRs). Discovered by PMAT-698i's [FWD-CACHE]
diagnostic logging.
The `warm!` macro in pre_warm_for_model:
macro_rules! warm {
($key:expr, $kernel:expr) => {{
let ptx = $kernel.emit_ptx_for_target(&target);
self.get_or_compile("silu_forward", &ptx)?; // <-- HARDCODED
count += 1;
}};
}
Every single `warm!()` call stored its compiled module under the
hashmap key "silu_forward", colliding on the first call:
1. warm!("batched_rmsnorm_fwd_896", BatchedVectorizedRmsNormKernel...)
→ cache["silu_forward"] = BatchedVectorizedRmsNorm PTX
2. warm!("gemm_forward_...", ...)
→ cache["silu_forward"] already Occupied → returns existing entry,
new PTX silently discarded
3-23. same — all subsequent kernels never actually pre-warm.
At runtime, every kernel looks up its real cache key:
let key = format!("batched_rmsnorm_fwd_{hidden_size}_eps{eps_bits:08x}");
match cache.get_cached(&key) { Some(m) => m, None => JIT }
— and cache-MISSES because the cache contains exactly one entry
under "silu_forward". JIT fires for every "pre-warmed" kernel during
the first forward pass — exactly when Blackwell sm_121's CUDA driver
crashes on cuModuleLoadData during active GPU work.
PMAT-698i's [FWD-CACHE] logging surfaced this: every kernel that was
"supposed to be pre-warmed" emitted [FWD-CACHE] Compiling at runtime,
proving the cache had nothing in it under those keys.
Fix: pass $key through to get_or_compile. One-character change
("silu_forward" → &key).
This explains the entire PMAT-698e..i cascade:
- PMAT-698e (workspace cap) — legit independent bug
- PMAT-698f (APR magic) — legit independent bug
- PMAT-698g (non-LoRA backward pre-warm) — would have been fine IF
forward pre-warm worked; the backward kernels were correctly stored
under their real keys (backward macro doesn't have the typo).
Defense-in-depth, still valuable.
- PMAT-698h (rms_norm_gamma_reduce) — same defense-in-depth.
- PMAT-698j (THIS) — the root cause.
The previous PMAT-698g/h fixes are still correct (they covered backward
gaps that exist independently). This PR addresses the forward cache,
which was the dominant source of post-pre-warm JIT events.
Test plan:
- [x] cargo check --features cuda — clean build
- [x] 366 autograd lib tests pass
- [ ] Live gx10 dispatch (post-merge) shows ZERO [FWD-CACHE] Compiling
events post-pre-warm (all 23 forward kernels now actually cached)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Merged
3 tasks
noahgift
added a commit
that referenced
this pull request
May 20, 2026
2026-05-20 — real distillation 1.5B teacher → 0.5B student on
Blackwell GB10 with the full PMAT-698e..n + PMAT-700-B cascade active.
initial_loss = 7.6746
final_loss = 7.2036 ← LESS THAN initial
62 steps, 122.7s, no errors
F-DISTILL-SMOKE-001 ("final_loss < initial_loss") discharged.
Phase 3 of SPEC-DISTILL-001 is COMPLETE.
Evidence:
- evidence/distill-phase-3-real-kd/dispatch.json — dispatch manifest
- evidence/distill-phase-3-real-kd/launch-final-pass.txt — full training log
Run dir on gx10: /home/noah/runs/distill-smoke-20260520-070404/
Trained student checkpoint: student-trained.apr/model.safetensors
Cascade summary (all merged):
- #1804 PMAT-700-B (cuBLAS prewarm skip)
- #1808 PMAT-698e (workspace cap)
- #1809 PMAT-698f (APR magic in weights loader)
- #1810 PMAT-698g (non-LoRA backward pre-warm)
- #1813 PMAT-698h (rms_norm_gamma_reduce pre-warm)
- #1815 PMAT-698i (FWD-CACHE diagnostic logging)
- #1817 PMAT-698j (THE root cause — warm! macro key)
- #1820 PMAT-698k (cache-key alignment: rope fwd + rmsnorm eps)
- #1823 PMAT-698m (smoke setup: non-degenerate batch)
- #1824 (post-mortem doc)
- #1827 PMAT-698n (rmsnorm pre-warm at both 1e-6 + 1e-5 eps)
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
THE root-cause bug
The
warm!macro inpre_warm_for_modelhadself.get_or_compile("silu_forward", &ptx)?hardcoded — every pre-warm call stored its compiled module under the SAME hashmap key. Only the first kernel actually got compiled; all subsequentwarm!()invocations hit the Occupied path and silently discarded their PTX.At runtime, every kernel looks up its real key (e.g.
batched_rmsnorm_fwd_896_eps358637bd) — cache miss → JIT mid-forward-pass → Blackwell sm_121 stream poisoning →forward_backward_with_grad returned None.Discovered via PMAT-698i diagnostic logging
The diagnostic PR's
[FWD-CACHE]log surfaced every "pre-warmed" kernel actually JIT-compiling at first runtime call. Proves the cache held exactly one entry (undersilu_forward), no matter how manywarm!()calls ran.Fix
macro_rules! warm { ($key:expr, $kernel:expr) => {{ + let key = $key; let ptx = $kernel.emit_ptx_for_target(&target); - self.get_or_compile("silu_forward", &ptx)?; + self.get_or_compile(&key, &ptx)?; count += 1; }}; }One-character behavioral change. The previous PMAT-698g/h fixes were defense-in-depth on the backward cache (whose
warm!macro was correct); this fixes the dominant forward-cache failure that all 5 prior iterations were chasing.Cascade summary
Test plan
cargo check --features cudaclean[FWD-CACHE] Compilingevents post-pre-warmfinal_loss < initial_loss(F-DISTILL-SMOKE-001)🤖 Generated with Claude Code