================================================================================================================ STEP 3C — FINAL EXACT Q1 TRAINING-PATH VALIDATION ================================================================================================================ Started UTC: 2026-07-20T23:12:02Z Repository: /content/Prism-llama.cpp Build directory: /content/Prism-llama.cpp/build Report: /content/prism_native_q1_lora/step03c_final_validation_report.txt ================================================================================================================ SOURCE PREFLIGHT ================================================================================================================ Resolved commit: 7529fdaaf99ffdc5ca71ace9c7409a56b27ad92f Expected commit: 7529fdaaf99ffdc5ca71ace9c7409a56b27ad92f [PASS] exact forward kernel [PASS] exact forward dispatch [PASS] exact backward-X kernel [PASS] native graph test [PASS] cross-backend tolerance ================================================================================================================ COMMENT-AWARE TRAINABILITY AUDIT ================================================================================================================ Raw regex matches, including comments: ['base_weight', 'x'] Executable ggml_set_param calls: ['x'] Expected executable calls: ['x'] [PASS] X is the only gradient probe. [PASS] Packed Q1 base is frozen. [PASS] Previous base_weight result came from a comment. ================================================================================================================ STRICT EXACT ORACLE ================================================================================================================ [PASS] Packed storage remained 1.125 bits/weight. [PASS] Exact forward matched the explicit CPU reference. [PASS] Exact backward-X matched the explicit CPU reference. [PASS] Finite differences matched analytical backward-X. ================================================================================================================ NATIVE FORWARD VALIDATION ================================================================================================================ [PASS] Run 1: /content/prism_native_q1_lora/step03b_validation/forward_01.log [PASS] Run 2: /content/prism_native_q1_lora/step03b_validation/forward_02.log [PASS] Run 3: /content/prism_native_q1_lora/step03b_validation/forward_03.log [PASS] Run 4: /content/prism_native_q1_lora/step03b_validation/forward_04.log [PASS] Run 5: /content/prism_native_q1_lora/step03b_validation/forward_05.log Native forward result: 5/5 PASS ================================================================================================================ NATIVE BACKWARD-X VALIDATION ================================================================================================================ [PASS] Native finite-difference test executed. [PASS] Exact packed-Q1 backward-X passed. [PASS] 17891/17891 gradient tests passed. [PASS] CUDA0 backend passed. ================================================================================================================ VALIDATION THRESHOLDS ================================================================================================================ Generic cross-backend forward NMSE limit: 0.0001000000 Finite-difference gradient MAA limit: 0.0020000000 [PASS] Gradient threshold was not relaxed. ================================================================================================================ SOURCE HASHES ================================================================================================================ 7cb8be48cc0c9101ba1097931522ae828cb3ce4797a603dcd380d3aac0e92a35 mulmat-q1-f32.cu 1f4e26bc75a368f3a9ac115788025ea8fba460348ffabdd2860f021c7729e9bc ggml-cuda.cu 11d87d709da74da094b7c086f8b8a2fa090de54e98e990566774cea0259362ed out-prod.cu f20a1ab6df7ef5ff9668214804e101836103de631890d6e674af7fc3ce0f9d90 test-backend-ops.cpp ================================================================================================================ CURRENT PATCH STATE ================================================================================================================ Git status: M ggml/src/ggml-cuda/ggml-cuda.cu M ggml/src/ggml-cuda/out-prod.cu M tests/test-backend-ops.cpp ?? ggml/src/ggml-cuda/mulmat-q1-f32.cu ?? ggml/src/ggml-cuda/mulmat-q1-f32.cuh Git diff stat: ggml/src/ggml-cuda/ggml-cuda.cu | 38 ++- ggml/src/ggml-cuda/out-prod.cu | 602 ++++++++++++++++++++++++++++++++++++---- tests/test-backend-ops.cpp | 124 +++++++++ 3 files changed, 703 insertions(+), 61 deletions(-) Status command return code: 0 Diff command return code: 0 ================================================================================================================ FINAL RESULT ================================================================================================================ [PASS] Exact packed-Q1 CUDA forward integrated. [PASS] Exact packed-Q1 CUDA backward-X integrated. [PASS] Default inference MMQ path remains separate. [PASS] Strict explicit forward oracle passed. [PASS] Strict explicit backward-X oracle passed. [PASS] Strict finite-difference oracle passed. [PASS] Native GGML forward passed 5/5 times. [PASS] Native GGML backward-X passed. [PASS] 17891/17891 gradient tests passed. [PASS] X is the only gradient probe. [PASS] Packed Q1 base is frozen. [PASS] No optimizer state exists for the packed base. [PASS] No persistent expanded weight matrix exists. Status: PASS Elapsed: 00:00 Step 3 is complete. The next stage is a native one-layer Q1 + LoRA optimizer-step test where only LoRA A/B receive gradients and AdamW state.