gitadmin/foxhunt
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

plan(dqn-v2): Plan 4 Task 2 GRN — decomposed into 2a/2b/2c with reality checks

Browse Source 
After a first dispatch attempt of monolithic Task 2 (GRN ADOPT) was
correctly refused with a thorough scope assessment, this revision
records the decomposition and the structural facts that drove it:

1. layout_fingerprint_seed() only fingerprints ISV slots, not
   param-tensor layout. The "fingerprint will auto-update" assumption
   in the original Task 2 spec was wrong for param-tensor reshuffles.
2. compute_param_sizes has 86 tensors (docstring saying 42 is stale).
   Inserting GRN's 9 sub-tensors shifts 82 downstream tensors and
   98 padded_byte_offset call sites.
3. At least 12 kernels consume h_s2 expecting ReLU non-negativity.
   GRN's LayerNorm output is zero-mean (signed). Per-consumer
   verification needed before swap.
4. crates/ml-supervised::tft::gated_residual is incompatible as a
   port: different formula (sigmoid gate, not GLU split) and
   incompatible tensor abstraction. New CUDA kernels from scratch.

Decomposition:
- Task 2a (research, no code): audit h_s2 consumers for ReLU vs
  LayerNorm semantics. Output: per-consumer table.
- Task 2b (small, checkpoint break): extend fingerprint seed to
  include param-tensor names + sizes. Pearl-aligned: complete
  fingerprint coverage rather than partial.
- Task 2c (large, checkpoint break): GRN kernels + 98 call-site
  migration + h_s2 consumer shims (per 2a). Blocked on 2a+2b.

Recommended order updated to thread 2a → 2b → 2c. Tasks 1, 6, 3
remain independent of 2c and can land in parallel where useful.
Pearl rules section added documenting the safety constraints
applied throughout the plan.

No code changes. Plan-doc only.
This commit is contained in:
jgrusewski
2026-04-25 11:11:13 +02:00
parent c247d9cab0
commit c0d3d7b2ff
1 changed files with 62 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
docs/superpowers/plans/2026-04-24-dqn-v2-plan-4-supervised-concepts.md
View File

@@ -125,15 +125,24 @@ digraph plan4_deps {
}
```
**Recommended order:**
1. **Task 5** — ISV-only attention diagnostics (light, no checkpoint break, gives telemetry for the heavier tasks)
2. **Task 4** — Encoder/decoder Rust-API split (refactor only, no kernel changes, no checkpoint break — sets clean boundaries for 1/3/6)
3. **Task 2** — GRN ADOPT (real architectural change, but localized to the 2 trunk Linear blocks; checkpoint break)
4. **Task 1** — Full VSN over feature groups (depends on E/D boundary; checkpoint break)
5. **Task 6** — Aux heads (depends on E/D boundary; checkpoint break)
6. **Task 3** — Multi-quantile IQN (most invasive — touches all IQN consumers; checkpoint break)
7. **Task 7** — Audit close-out (pure docs, after 1-6)
8. **Task 8** — Argo validation run (after 1-7 land)
**Recommended order (revised 2026-04-25 after Task 2 decomposition):**
1. **Task 5 Mode A** — ISV-only attention diagnostics — ✅ landed `cfc4ccb72`
2. **Task 4** — Encoder/decoder Rust-API split — ✅ landed `fbc299fa2`
3. **Legacy CPU-DtoH cleanup** — `per_branch_vsn_mean` + `per_branch_target_drift` deleted — ✅ landed `c247d9cab`
4. **Task 2a** — Audit h_s2 consumers for ReLU vs LayerNorm semantics (research, no code) — **next up**
5. **Task 2b** — Extend layout_fingerprint_seed to cover param-tensor layout (small commit, checkpoint break)
6. **Task 2c** — GRN ADOPT proper (large commit, checkpoint break) — blocked on 2a + 2b
7. **Task 1** — Full VSN over feature groups (depends on 2c trunk shape; checkpoint break)
8. **Task 6** — Aux heads (independent of 2c — can land in parallel; checkpoint break)
9. **Task 3** — Multi-quantile IQN (independent of trunk; checkpoint break)
10. **Task 7** — Audit close-out (pure docs, after 1/2c/3/6)
11. **Task 8** — Argo validation run (terminal)
**Pearl-derived safety rules applied throughout:**
- `pearl_cold_path_no_exception_to_gpu_drives.md` — every producer (mean, EMA, RMS, etc.) goes through a GPU kernel even on cold path. No CPU DtoH+loop helpers.
- `pearl_one_unbounded_signal_per_reward.md` — exactly one unbounded multiplicand per reward term; bound everything else to [0,1].
- `feedback_no_partial_refactor.md` — shared-contract changes (param-tensor layout, ISV layout) migrate every consumer in the same commit.
- `feedback_no_legacy_aliases.md` — no compat shims for renamed/removed code paths.
**Checkpoint break consolidation:** Tasks 2, 1, 6, 3 each break checkpoint compatibility. Land them in sequence with NO Argo run between (one fingerprint shift per commit; final Argo run at Task 8 amortizes the retraining cost across all changes). Revert to a known-good commit if any single task regresses the local smoke beyond ~25-point Sharpe drop.
@@ -383,9 +392,51 @@ git commit -m "plan4(task1C): upgrade E.1 VSN to per-feature weights (cond. on 1
---
## Task 2: E.2 Gated Residual Network — Branch A (ADOPT) confirmed
## Task 2: E.2 Gated Residual Network — Branch A (ADOPT), DECOMPOSED into 2a/2b/2c
**Reality reconciliation (2026-04-25).** GRN audit completed: GRN exists in `crates/ml-supervised/src/tft/gated_residual.rs` (Rust + supervised pretraining only) and `crates/ml/src/transformers/temporal_fusion.rs` (TFT inference path). It is **NOT** wired into the DQN cuda_pipeline trunk. The DQN trunk in `batched_forward.rs::forward_online_raw` is `Linear→ReLU→Linear` with no LayerNorm, no ELU, no GLU. **Branch A (ADOPT) applies.** Branch B is moot.
**Reality reconciliation (2026-04-25, second iteration).** GRN audit completed: GRN exists in `crates/ml-supervised/src/tft/gated_residual.rs` (Rust + supervised pretraining only) and `crates/ml/src/transformers/temporal_fusion.rs` (TFT inference path). It is **NOT** wired into the DQN cuda_pipeline trunk. The DQN trunk in `batched_forward.rs::forward_online_raw` is `Linear→ReLU→Linear` with no LayerNorm, no ELU, no GLU. **Branch A (ADOPT) applies.** Branch B is moot.
### Why this decomposed (post-2026-04-25 dispatch attempt)
A first dispatch of monolithic Task 2 was refused by the implementer with a thorough scope assessment. Three structural facts changed the calculus:
1. **`layout_fingerprint_seed()` only covers ISV slots, not param-tensor layout.** The current scheme can detect ISV-shape changes but param-tensor reshuffles (which GRN will cause) pass silently through checkpoint load. The "fingerprint will auto-update" assumption was wrong.
2. **`compute_param_sizes` has 86 tensors** (the docstring saying 42 is stale). Inserting GRN's 9 sub-tensors at trunk position shifts 82 downstream tensors and **98 `padded_byte_offset` call sites** across the cuda_pipeline. Single-commit migration is required (per `feedback_no_partial_refactor.md`) but per-site verification is non-trivial.
3. **At least 12 kernels consume `h_s2`** (VSN gates, Mamba2 history, attention, plan head, KAN spline, ISV feature gate, ISV temporal route, regime dropout, prediction loss, selectivity head, conf head, value/branch decoders). All currently see `ReLU(...)` output (non-negative); GRN's LayerNorm output is zero-mean and contains negatives. Whether each consumer's downstream arithmetic tolerates negative inputs is unverified.
The existing supervised GRN at `crates/ml-supervised::tft::gated_residual` is **incompatible** as a port: different formula (sigmoid gate, not GLU split) and different tensor abstraction (no cuBLAS). New CUDA kernels must be written from scratch.
### Decomposition
#### Task 2a: Audit h_s2 consumers for ReLU vs LayerNorm semantics (research only, no code changes)
Inventory every kernel/site that reads the trunk's `h_s2` buffer. For each: (i) document what arithmetic the consumer applies, (ii) flag whether it implicitly assumes `h_s2 ≥ 0`, (iii) propose mitigation if so (e.g., `relu(h_s2)` shim at the consumption site, or design change to handle signed input). Output: a table appended to `docs/ml-supervised-to-dqn-concept-audit.md` (or a new audit doc), one row per consumer.
This is the unblock for Task 2c — without per-consumer signoff, the trunk swap risks silent breakage.
#### Task 2b: Extend `layout_fingerprint_seed()` to cover param-tensor layout (small commit, checkpoint break)
The current seed string fingerprints ISV slot names + indices. Extend it to also include param-tensor names + sizes (e.g., `W_S1=86,SH1xSD;b_S1=87,SH1;…`). Re-bake `LAYOUT_FINGERPRINT_CURRENT`. Add the param-layout assertion to checkpoint load alongside the ISV fingerprint check. With this in place, **any** param-tensor reshuffle will trigger fail-fast on stale checkpoints — including Task 2c's GRN insertion.
This task itself is a one-commit checkpoint break (the seed string changes, fingerprint changes, all current checkpoints fail load — but no model behavior changes, so the smoke is just "does training still complete from cold start").
Pearl: same family as `pearl_cold_path_no_exception_to_gpu_drives.md` — the fingerprint contract should be **complete** for what it claims to protect, not partial. Partial coverage is worse than no coverage because it implies safety where there isn't.
#### Task 2c: GRN ADOPT proper (large commit, checkpoint break)
With 2a's per-consumer signoff and 2b's structural fingerprint in place, the actual ADOPT becomes mechanical:
- New `grn_kernel.cu` (forward + backward, composed from cuBLAS GEMM + small custom kernels: `grn_elu_inplace`, `grn_glu_split`, `grn_layernorm_forward`, `grn_layernorm_backward`, `grn_residual_add`)
- New `gpu_grn.rs` (Rust wrapper, Adam state, init paths)
- Modify `compute_param_sizes`: insert GRN_h_s1 (7 tensors: W_a, b_a, W_b, b_b, W_residual, γ, β) and GRN_h_s2 (6 tensors: same minus W_residual since SH1==SH2). Total +9 trunk tensors.
- Migrate **all 98 `padded_byte_offset` call sites** that index ≥ trunk-end position (per `feedback_no_partial_refactor.md`)
- Update `batched_forward.rs::encoder_forward_only` and `batched_backward.rs` trunk path
- Apply 2a-derived shims at h_s2 consumers if any required non-negativity
- Update audit docs
**No CPU compute introduced.** All GRN ops on GPU (cuBLAS GEMMs + custom element-wise kernels). The pearl `pearl_cold_path_no_exception_to_gpu_drives.md` applies: even diagnostic-side reductions of GRN params (if any) go through the same multi-block GPU kernel pattern Tasks 5 and the legacy-cleanup follow-up established.
### Original task body (pre-decomposition reference, kept for design context)
**Implementation surface (concrete):**
- **Trunk site**: `batched_forward.rs::forward_online_raw` lines 471–481 (h_s1 + h_s2 GEMMs). Replace each `Linear→ReLU→Linear` block with a GRN block.