Repository navigation
Conversation
specs/igla/coder/preflight.t27 decides GO or REFUSE (reason code) before a run is launched: shape checks (batch, heads, kv heads), planned hours against the queue timeout, planned cost against the cap. specs/igla/coder/kill_rule.t27 decides CONTINUE, WARN or KILL at an eval point: non-finite numbers, a grad_norm spike, or val_bpb above 2.5x a curve scaled from the 100m ternary run. Tests use real igla-coder-gpu numbers. 13/13 and 12/12 pass on the Railway lab, 0 vacuous; both are sealed; the gen-c output compiles clean and its own test runner passes. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
gHashTag
added
the
no-auto-merge
Contributor
This was referenced Oct 18, 2026
Closed
Merged
Merged
Merged
Merged
This was referenced Oct 18, 2026
Merged
Merged
Merged
Merged
Contributor
PR DashboardGenerated at: 2026-10-18 16:46:28 UTC
Summary
Seal Status
|
This branch has not been deployed
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Closes #10363
What
Two decision specs for the igla-coder-gpu training pipeline. The PyTorch kernels stay Python (t27c has no GPU backend); the decisions around a run are plain logic and live here.
specs/igla/coder/preflight.t27-- GO or REFUSE with a reason code, before a run is launched. Order: unusable input (6), batch not a multiple of micro x seq_len (3), d_model % n_head (4), n_head % n_kv_head (5), planned hours >= queue timeout (1), planned cost > cap (2). The speed comes from the 60-step smoke run, derated by 0.95 (the real 4090 run took 19.74 h against 19.49 h predicted from the smoke speed).specs/igla/coder/kill_rule.t27-- CONTINUE, WARN or KILL at an eval point. A non-finite loss, grad_norm or val_bpb kills at any eval. A grad_norm above 100, or a val_bpb above 2.5x the reference curve, warns on the first eval and kills from the second. Nothing but non-finite is judged below 500M tokens. The curve is 0.4722 bpb at 10B tokens (the 100m ternary run) scaled by (10B/T)^(5/64), which needs only square roots; it gives 0.5262 at 2.5B against the real 0.5248..trinity/seals/coder_igla-coder-*.json).Evidence (Railway lab, never the Mac)
t27c test-report: preflight 13/13, kill_rule 12/12, 0 vacuous passes.t27c gen-cof both specs compiles with 0 errors and 0 warnings (cc -shared -fPIC ... -lm), and the generated test runner (-DT27_TEST_MAIN) prints "All 13 tests passed." and "All 12 tests passed.".Honest limits
runs/is not committed). The kill rule is calibrated on the ladder and sweep endpoints, the HF final summaries and the fine-tune log.healthy_trace_has_no_killsuses an ASSUMED plateau factor of 1.3 over the curve and says so. Once a real per-eval 100m log exists, add it as a test.t27c gen-python/gen-jslower declarations, not bodies (by design,codegen_python.rs), so they emit only the constants of these specs.gen-clowers bodies, so the igla-coder-gpu glue builds the vendored C withccand loads it with ctypes.gen-cdoes not lower the@absbuiltin (the# t27c gen-coutput has a stray@); the specs avoid it.Glue and vendored output are in the matching gHashTag/igla-coder-gpu PR (gHashTag/igla-coder-gpu#44, which resolves igla-coder-gpu issue 43 there). The running tern500m2 pod and the epic's LEDGER entry are untouched.
The PR carries
no-auto-merge: the specs should be read before they gate real GPU spend.🤖 Generated with Claude Code