feat: E_PT2 energy correction for HI-NQS (Issue #35 Tier 1 #2)#37
Draft
thc1006 wants to merge 10 commits intoQuantumNoLab:mainfrom
Draft
feat: E_PT2 energy correction for HI-NQS (Issue #35 Tier 1 #2)#37thc1006 wants to merge 10 commits intoQuantumNoLab:mainfrom
thc1006 wants to merge 10 commits intoQuantumNoLab:mainfrom
Conversation
… Tier 1 #2) E_PT2 = Σ_{x∉V} ⟨x|H|Ψ₀⟩² / (E₀ - H_xx) Standalone pure function in _pt2_helpers.py with 5 TDD tests: - returns float, negative from HF reference, corrected energy closer to FCI, full basis gives zero
- Add compute_pt2_correction field to HINQSSQDConfig (default False) - When enabled, metadata includes e_pt2 and corrected_energy - Size guard prevents crash when coeffs/basis mismatch - 43 tests passing (8 new for E_PT2) Co-authored-by: leo07010 <leo07010@gmail.com>
Tests use mock hamiltonian without integrals; IBM solver auto-detect tries to access hamiltonian.integrals.h1e causing AttributeError.
- compute_e_pt2 now receives cumulative_basis + full-basis eigenvector (not prev_batch_configs which is a random sub-sample) - E₀ for PT2 denominator comes from same diag as Ψ₀ (consistency) - Removed unused external_hxx dict in compute_e_pt2
- Fix docstring: |⟨x|H|Ψ₀⟩|² (abs-squared), clarify real-valued note - Emphasize basis must be FULL variational space in docstring - Remove unused basis_coeff_map from compute_e_pt2 - Add pt2_e0 and pt2_wall_time to metadata for traceability - corrected_energy uses pt2_e0 (from full-basis diag), not best_energy
- test_metadata_contains_pt2_e0_and_wall_time: verify traceability - test_corrected_energy_uses_pt2_e0: verify corrected = pt2_e0 + e_pt2 (not best_energy + e_pt2)
- CHANGELOG: E_PT2 correction, compute_pt2_correction config, tri-state use_ibm_solver - api_reference: compute_e_pt2 function signature and description
There was a problem hiding this comment.
Pull request overview
Adds optional Epstein–Nesbet PT2 (EN-PT2) energy correction to the HI+NQS+SQD workflow to produce a post-variational corrected energy, along with unit/integration tests to validate the correction behavior and metadata wiring.
Changes:
- Introduced
compute_e_pt2()to compute EN-PT2 energy corrections over external determinants. - Added
compute_pt2_correction(opt-in) toHINQSSQDConfigand integrated PT2 correction + metadata intorun_hi_nqs_sqd. - Expanded tests for
compute_e_pt2and end-to-end PT2 integration; patched IBM SQD availability in initial-basis tests.
Reviewed changes
Copilot reviewed 6 out of 6 changed files in this pull request and generated 4 comments.
| File | Description |
|---|---|
src/qvartools/methods/nqs/_pt2_helpers.py |
Adds compute_e_pt2() implementation for EN-PT2 correction. |
src/qvartools/methods/nqs/hi_nqs_sqd.py |
Adds config flag + optional full-basis diag and PT2 correction metadata wiring. |
tests/test_methods/test_pt2_selection.py |
Adds unit tests for compute_e_pt2 and integration tests for PT2 correction in run_hi_nqs_sqd. |
tests/test_methods/test_initial_basis.py |
Patches _IBM_SQD_AVAILABLE to stabilize tests when mocking the solver. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
1. Docstring: corrected_energy = pt2_e0 + e_pt2 (not energy + e_pt2) 2. Dense diag threshold: 50K → 8K (align with CIPSI _SPARSE_DIAG_THRESHOLD) 3. Memory leak: connections[j].detach().clone() to avoid retaining full tensor 4. Intruder state: clamp near-zero denominator instead of skipping
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
1. Raise ValueError if coeffs.shape[0] != basis.shape[0] (fail-fast) 2. Use diagonal_elements_batch() instead of per-config diagonal_element() (faster + frees config tensors after batched computation)
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
1. Docstring: add diagonal_elements_batch to required Hamiltonian API 2. Store external configs on CPU (.detach().cpu().clone()) to reduce GPU OOM risk 3. Chunked diagonal computation (batch_size=4096) to bound peak memory
Collaborator
Code reviewNo issues found. Checked for bugs and CLAUDE.md compliance. 🤖 Generated with Claude Code |
Member
Author
|
張人魚說這先不要合併 |
Collaborator
|
OK |
thc1006
added a commit
to thc1006/qvartools
that referenced
this pull request
Apr 7, 2026
Adds the four NQS methods from src/qvartools/methods/nqs/ as first-class runnable pipeline catalog entries, alongside the existing 001-009 ablation groups. Follows the convention "one method = one folder, variants = multiple .py inside the folder, one multi-section YAML per method". New pipelines (7 scripts in 4 folders): - 010_hi_nqs_sqd/ (HI+NQS+SQD, iterative, optional PT2/IBM) * default.py — baseline, auto IBM detect * pt2.py — use_pt2_selection=True + temperature annealing * ibm_off.py — use_ibm_solver=False (force GPU fallback) - 011_hi_nqs_skqd/ (HI+NQS+SKQD, iterative + Krylov expansion) * default.py — baseline (use_ibm_solver=False) * ibm_on.py — use_ibm_solver=True + S-CORE recovery - 012_nqs_sqd/ (two-stage NQS+SQD, no iteration) * default.py - 013_nqs_skqd/ (two-stage NQS+SKQD with Krylov expansion) * default.py Each YAML has one section per variant. Each wrapper script is ~80 LoC, uses the existing experiments/config_loader.py for CLI+YAML, calls the qvartools.methods.nqs.run_*() function, and prints results in the format that run_all_pipelines.py's regex parsers expect (Best energy, Final energy, Wall time). run_all_pipelines.py PIPELINES list extended from 26 to 33 entries. Bonus fix discovered during smoke testing: nqs_sqd.py and nqs_skqd.py used mol_info["n_orbitals"] directly, but get_molecule() does NOT populate that key — it must come from hamiltonian.integrals. The HI methods avoided this via extract_orbital_counts(); the simple methods did not. Fixed by routing both simple methods through extract_orbital_counts() too. This is a pre-existing bug that nothing was exercising before; now fixed and verified. Smoke tests on H2 (CPU): - 010 hi_nqs_sqd/default: E=-1.1372838345 Ha (0.0 mHa error) 10.5s - 011 hi_nqs_skqd/default: E=-1.1372838345 Ha (0.0 mHa error) 10.6s - 012 nqs_sqd/default: E=-1.1167593074 Ha (20.5 mHa error) 74.6s (HF, expected — pure NQS ablation) - 013 nqs_skqd/default: E=-1.1372838345 Ha (0.0 mHa error) 44.4s All 52 tests in tests/test_methods/ + tests/test_run_all_pipelines.py pass. ruff check + ruff format clean. Note: 010_hi_nqs_sqd/correction.py (compute_pt2_correction=True variant) is intentionally deferred — that config field only exists on the feat/e-pt2-correction branch (PR QuantumNoLab#37). Will add as a follow-up commit after PR QuantumNoLab#37 merges to main and this branch rebases.
8 tasks
thc1006
added a commit
that referenced
this pull request
Apr 15, 2026
…peline entries (#39) * docs(adr-003): cross-reference Issue #38 and correct amd-sbd speedup - Added "Update 2026-04-07" section noting that AMD-HPC/amd-sbd's 95x is measured on MI250X (AMD GPU), not GB200 (which is only 2.64x in the same paper). This significantly weakens the case for the C++/sbd path on NVIDIA hardware. - Added cross-references to Issue #38 (CuPy factored-space Davidson, parallel path), and to memory/project_phase2b_davidson_commitment.md. - Documented that Phase 2b proper Davidson is non-negotiable for multireference chemistry (Cr2, dissociation, open-shell) regardless of which backend wins. * refactor(pipelines): rename folders 01-09 to 001-009 (3-digit prefix) Renames the 9 existing pipeline group folders and their YAML configs from 2-digit to 3-digit prefix to leave room for the upcoming 010+ method-as-pipeline catalog (Issue: standardized open-source benchmark catalog). Changes: - git mv all 9 pipeline folders: 01_dci → 001_dci, ..., 09_vqe → 009_vqe - git mv all 9 YAML configs to match - Update first-line comment in each YAML to use new prefix - Update PIPELINES list in run_all_pipelines.py (26 entries) - Update --skip-iterative check from ("06","07","08") to ("006","007","008") - Update --skip-quantum check from "09_vqe" to "009_vqe" - Update --only help text and docstring example to 3-digit - Update tests/test_run_all_pipelines.py (4 references to 09_vqe) This is a pure rename + reference update. No behaviour change. All 3 tests in test_run_all_pipelines.py still pass. Doc updates (AGENTS.md, README.md, RST files) are scheduled for Phase 4 of this refactor in a separate commit. * fix: patch _IBM_SQD_AVAILABLE in initial_basis tests Tests use mock hamiltonian without integrals; IBM solver auto-detect tries to access hamiltonian.integrals.h1e causing AttributeError. * refactor(methods/nqs): extract shared helpers and add METHODS_REGISTRY Light additive refactor of the four NQS methods to remove duplication and expose a registry that the upcoming 010+ pipeline wrappers (and any future benchmark catalog tooling) can use to dispatch by method id. Changes: - Add src/qvartools/methods/nqs/_shared.py with three helpers extracted from the four method runners (build_autoregressive_nqs, extract_orbital_counts, validate_initial_basis). Bit-equivalent to the inlined code that was there before. - Update nqs_sqd.py, nqs_skqd.py, hi_nqs_sqd.py, hi_nqs_skqd.py to use the new helpers. No behaviour change, no API change, no signature change. - Add METHODS_REGISTRY dict to methods/nqs/__init__.py mapping {nqs_sqd, nqs_skqd, hi_nqs_sqd, hi_nqs_skqd} -> run_fn, config_cls, metadata flags (iterative/has_krylov_expansion/has_ibm_solver/ has_pt2_selection/supports_initial_basis), description, and pipeline_folder. Verified: - All 49 tests in tests/test_methods/ pass (was 46 + 3 cherry-picked failures from main; cherry-pick of 3f39fe2 fixed the 3). - 62 tests across test_run_all_pipelines + test_methods + test_nqs/ test_adapters all pass. - Zero changes to algorithmic logic, exception types, return shapes, or function signatures. * feat(pipelines): add 010-013 method-as-pipeline catalog (NQS methods) Adds the four NQS methods from src/qvartools/methods/nqs/ as first-class runnable pipeline catalog entries, alongside the existing 001-009 ablation groups. Follows the convention "one method = one folder, variants = multiple .py inside the folder, one multi-section YAML per method". New pipelines (7 scripts in 4 folders): - 010_hi_nqs_sqd/ (HI+NQS+SQD, iterative, optional PT2/IBM) * default.py — baseline, auto IBM detect * pt2.py — use_pt2_selection=True + temperature annealing * ibm_off.py — use_ibm_solver=False (force GPU fallback) - 011_hi_nqs_skqd/ (HI+NQS+SKQD, iterative + Krylov expansion) * default.py — baseline (use_ibm_solver=False) * ibm_on.py — use_ibm_solver=True + S-CORE recovery - 012_nqs_sqd/ (two-stage NQS+SQD, no iteration) * default.py - 013_nqs_skqd/ (two-stage NQS+SKQD with Krylov expansion) * default.py Each YAML has one section per variant. Each wrapper script is ~80 LoC, uses the existing experiments/config_loader.py for CLI+YAML, calls the qvartools.methods.nqs.run_*() function, and prints results in the format that run_all_pipelines.py's regex parsers expect (Best energy, Final energy, Wall time). run_all_pipelines.py PIPELINES list extended from 26 to 33 entries. Bonus fix discovered during smoke testing: nqs_sqd.py and nqs_skqd.py used mol_info["n_orbitals"] directly, but get_molecule() does NOT populate that key — it must come from hamiltonian.integrals. The HI methods avoided this via extract_orbital_counts(); the simple methods did not. Fixed by routing both simple methods through extract_orbital_counts() too. This is a pre-existing bug that nothing was exercising before; now fixed and verified. Smoke tests on H2 (CPU): - 010 hi_nqs_sqd/default: E=-1.1372838345 Ha (0.0 mHa error) 10.5s - 011 hi_nqs_skqd/default: E=-1.1372838345 Ha (0.0 mHa error) 10.6s - 012 nqs_sqd/default: E=-1.1167593074 Ha (20.5 mHa error) 74.6s (HF, expected — pure NQS ablation) - 013 nqs_skqd/default: E=-1.1372838345 Ha (0.0 mHa error) 44.4s All 52 tests in tests/test_methods/ + tests/test_run_all_pipelines.py pass. ruff check + ruff format clean. Note: 010_hi_nqs_sqd/correction.py (compute_pt2_correction=True variant) is intentionally deferred — that config field only exists on the feat/e-pt2-correction branch (PR #37). Will add as a follow-up commit after PR #37 merges to main and this branch rebases. * docs(pipelines): update for 3-digit catalog and add 010-013 documentation Phase 4 of the pipeline catalog refactor: update every doc file that references the old 0X_ pipeline paths to use the new 3-digit names, and add a new section documenting the 010-013 method-as-pipeline catalog. Files updated: - AGENTS.md * directory tree: list all 13 pipeline folders (001-009 + 010-013) * config_loader pattern examples use 002_nf_dci paths and 010 wrappers * "24 Pipeline Variants" → "33 Pipeline Variants" - README.md * Quick start examples show both 001_dci and 010_hi_nqs_sqd usage * --only example uses 3-digit prefixes - docs/experiments_guide.md * All Group 01-09 headers renumbered to 001-009 * All 24 ablation script paths updated to new prefixes * --only example: 01 02 04 → 001 002 004 * "all 24 pipelines" → "all 33 pipelines" * NEW Tier 2 section documenting 010-013 method-as-pipeline catalog with per-group variant tables and the numbering convention - docs/source/user_guide/pipelines.rst * Restructured to show "Tier 1: ablation groups (001-009)" and "Tier 2: method-as-pipeline catalog (010-013)" * Added Tier 2 table with method ids, variants, and METHODS_REGISTRY note * Added "Numbering convention" subsection (001-009 / 010-099 / 100-199 / 200+) * Updated iterative pipeline code example (mol_info workaround removed — extract_orbital_counts handles it now) * CLI examples updated to 3-digit - docs/source/user_guide/yaml_configs.rst * Tier 1 / Tier 2 split for the YAML file table * Added 009_vqe.yaml + the 4 new 010-013 multi-section YAMLs * CLI examples updated to 3-digit - docs/source/tutorials/h2_pipeline.rst - docs/source/tutorials/beh2_pipeline.rst * CLI examples updated to 3-digit * h2 tutorial: added a 010-013 example - docs/decisions/002-eliminate-torch-numpy-roundtrips.md * Updated the validation script paths and "24/24" → "33/33" Verified: grep for 0[1-9]_(dci|nf_dci|nf_only|hf_only|iterative_nqs|vqe) in the live tree (excluding .claude/worktrees) returns zero matches. * fix: address 10 Copilot review findings on PR #39 Bug 1 — experiments/pipelines/run_all_pipelines.py:376 --skip-iterative was only skipping groups 006/007/008 but 010_hi_nqs_sqd and 011_hi_nqs_skqd are also iterative (HI methods). Added 010/011 to the prefix check. Verified: all 5 iterative groups now listed in skipped set when --skip-iterative is active. Bugs 2-8 — 7 wrapper scripts at line 56 (YAML section precedence) load_config() mutates args.device and args.molecule with merged defaults (line 125 of config_loader.py: setattr for every merged key), so the wrapper's `args.device or section.get("device")` short-circuited on args.device="auto" and silently ignored the YAML section's device value. Fixed by importing _get_explicit_cli_args from config_loader and only using the CLI arg when the user explicitly typed --device on the command line. Same fix applied to --molecule (positional). Verified: running with --device cpu correctly overrides section; running without --device correctly picks up the section value (auto→cuda detection). Nit 9 — src/qvartools/methods/nqs/__init__.py docstring Docstring claimed METHODS_REGISTRY was "used by the 010-013 wrapper scripts so they can dispatch by name without hard-importing each method module", but the wrappers were doing direct imports. Fixed architecturally: all 7 wrappers now dispatch via METHODS_REGISTRY[METHOD_KEY] and only import METHODS_REGISTRY (not individual config/run pairs). The __init__.py itself still imports all four methods because the registry needs callable references — that's a Python fundamental, not a design flaw. Nit 10 — src/qvartools/methods/nqs/_shared.py docstring Docstring said "behaviour-preserving extractions only — no API changes", but extract_orbital_counts adds the mol_info→hamiltonian.integrals fallback to nqs_sqd/nqs_skqd, which fixes a pre-existing bug in those runners (they were end-to-end broken because they accessed mol_info["n_orbitals"] directly but get_molecule doesn't populate it). Updated docstring to distinguish the three scopes: (1) build_autoregressive_nqs = pure refactor, (2) validate_initial_basis = pure refactor, (3) extract_orbital_counts = pure refactor for HI methods + bug fix for simple methods. Testing: - 72 tests pass (test_methods + test_run_all_pipelines + test_config_loader) — up from 62 in the previous commit (test_config_loader added to the test surface this round) - ruff check + ruff format clean - Smoke tested 010_hi_nqs_sqd/default.py h2: reaches FCI exactly (0.0 mHa) - Smoke tested 012_nqs_sqd/default.py h2 both with --device cpu (CLI override works) and without --device (section default picked up, auto→cuda detection) - --skip-iterative correctly lists all 5 iterative groups (006,007,008,010,011) in the skipped set Out of scope (pre-existing, not caused by this PR): - 007_iterative_nqs_dci/iter_nqs_dci_sqd.py and iter_nqs_dci_krylov_classical.py fail with "initial_basis must be integer or bool dtype ... got torch.float32". Verified these same scripts fail identically on main (stash-checkout-unstash test). Pre-existing bug in the 007 group's handling of initial_basis. Should be filed as a separate issue; not touched in this PR. * refactor+docs: address Copilot round-2 finding + self-review P0/P1/P2 items Option C: fix all 8 items from my own deep review on PR #39 plus Copilot's one remaining round-2 finding, then re-ran the full smoke + lint + test net and did a second ultrathink code review before commit. 1. docs/source/user_guide/pipelines.rst (Copilot round 2) Tier 2 table rows now use the actual on-disk folder names with underscore and monospace formatting (``010_hi_nqs_sqd`` etc.), not the human-facing "010 hi_nqs_sqd" form. Added a new "Programmatic dispatch via ``METHODS_REGISTRY``" subsection documenting all 9 registry fields + an idiomatic code example. Rewrote the Iterative Pipelines example to show the registry dispatch pattern (matches how the 010-013 wrappers now work) while noting that direct imports still work. 2. CHANGELOG.md (P0: user-visible breaking changes were undocumented) Added entries under Changed (2x BREAKING for rename + --only), Added (Tier 2 catalog, METHODS_REGISTRY, _shared.py, public get_explicit_cli_args), and Fixed (nqs_sqd/nqs_skqd mol_info bug rediscovered via smoke test). 3. experiments/pipelines/run_all_pipelines.py (P0: --only silent failure) Detect pre-2026-04-07 2-digit --only format and print a migration warning to stderr with the recommended 3-digit form. "--only 01 02" now emits "use --only 001 002 instead" instead of silently running 0 pipelines. Verified at runtime. 4. experiments/config_loader.py (P1: private symbol imported externally) Renamed _get_explicit_cli_args -> get_explicit_cli_args (public name is now primary). Added _get_explicit_cli_args as a backward-compat alias at module bottom so tests/test_config_loader.py keeps working without changes. Docstring updated to explain the alias. Internal call in load_config() uses the public name. 5-6. 7 wrapper scripts in 010-013 (P1: multi-section YAML silent fallback) Regenerated via template with two new module-level helpers: - _resolve_section(cfg, variant): extracts the requested section from multi-section YAML, warns when falling back to flat cfg if OTHER sections exist (indicating the user picked the wrong script or the wrong --config file) - _build_config_kwargs(section, config_cls, device): filters section keys to dataclass fields, warns on unknown keys (catches YAML typos like "n_itrations" silently being dropped) Also switched to importing get_explicit_cli_args (public name). Verified at runtime: wrong-section fallback warns, typo'd YAML key warns, default path still reaches FCI exactly on H2. 7. docs/source/user_guide/pipelines.rst (P2: METHODS_REGISTRY undocumented) Covered by #1 above — new subsection documents the registry. 8. docs/source/user_guide/pipelines.rst (P2: iterative example inconsistent with wrapper pattern) Covered by #1 above — iterative example now uses registry dispatch. Verification after all 8 fixes: - 61 tests pass (tests/test_methods/ + tests/test_run_all_pipelines.py + tests/test_config_loader.py -m "not slow and not gpu and not pyscf") - ruff check src/ tests/ experiments/ — clean - ruff format --check src/ tests/ experiments/ — 206 files already formatted - Runtime smoke tests on H2/CPU: * --only 01 02 emits migration warning (verified) * multi-section YAML with missing section emits fallback warning (verified) * YAML with typo'd key emits unknown-key warning (verified) * 010_hi_nqs_sqd/default.py h2 still reaches FCI (0.0 mHa, ~9s) * run_all_pipelines.py --only 001 005 012 013 --skip-quantum: 6/6 pass - All 7 wrappers import cleanly with correct METHOD_KEY/VARIANT constants - nqs_sqd/nqs_skqd still use extract_orbital_counts (grep confirmed) - RST heading "Programmatic dispatch via ``METHODS_REGISTRY``" is 46 chars, underline is 46 ^ (Sphinx requires >= title length, exact match OK) - METHODS_REGISTRY 9 fields enumerated in docs match __init__.py definition Second code review findings (all minor, none blocking): - _resolve_section + _build_config_kwargs helpers duplicated across 7 wrapper files (~280 lines). Could extract to shared _wrapper_helpers.py module in a follow-up. Not a bug. - --only warning does not catch single-digit --only 1 2. Low impact. - _resolve_section's isinstance(v, dict) filter could miss dict-valued config fields. No current NQS config has dict fields, so no collision. - CHANGELOG get_explicit_cli_args entry re-worded to clarify the public name is now primary, not the other way around.
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.
feat: E_PT2 EN-PT2 能量修正（Issue #35 Tier 1 #2）
變更內容
compute_e_pt2：新增 EN-PT2 能量修正函式（_pt2_helpers.py）E_PT2 = Σ_{x∉V} |⟨x|H|Ψ₀⟩|² / (E₀ - H_xx)HINQSSQDConfig.compute_pt2_correction：新增 opt-in 欄位（defaultFalse）run_hi_nqs_sqd：啟用時在最終 iteration 後做 full-basis diag + E_PT2 計算e_pt2、pt2_e0、corrected_energy、pt2_wall_timeuse_ibm_solver改為 tri-state：None=auto、True=強制啟用、False=強制停用_SPARSE_DIAG_THRESHOLD：10K → 8K（default params 下可觸發 sparse path）test_initial_basis.py：patch_IBM_SQD_AVAILABLE避免 mock hamiltonian crashBenchmark 結果（CPU，Numba + IBM solver）
E_PT2 修正量較大（135-190 mHa），表示 NQS basis 仍不完整。結合 Tier 1 #1（H-connection expansion）可大幅縮小。
GPU 對比（DGX Spark GB10）
結論：當前架構下 CPU 更快。GPU 加速需替換 IBM solver（Issue #35 Tier 3）。
下一步加速路線（調研結果）
TDD 測試
Checklist
Co-authored-by: leo07010 leo07010@gmail.com