llama.cpp-sm70-cuda12.9 / REVALIDATION.json
orwelian84's picture
llama.cpp f74e2fa for NVIDIA Volta (sm_70), CUDA 12.9 — released 2026-10-20 with the Q5_1 caveat stated
0b864c5 verified
Raw History Blame Contribute Delete
15.7 kB
{
"schema": 1,
"updated_utc": "2026-10-20T20:43:34.031803+00:00",
"source_commit": "f74e2fa6a31cc33ab4ee4ae516641cfe62df3924",
"modified_source": true,
"passed": false,
"release_ready": false,
"verdict": "Fresh numerical tests pass; release blocked by the original uncaptured Q5_1 input under the every-failure evidence requirement.",
"gates": {
"kernels": {
"passed": false,
"fresh_runs_passed": true,
"historical_failure_evidence_complete": false
},
"model_agreement": {
"passed": true
},
"flash_attention": {
"passed": true,
"targeted_regression_passed": 1,
"full_family_passed": 12012,
"full_family_failed": 0,
"full_family_unsupported": 4017
},
"package": {
"passed": true,
"binary_relocation_passed": true,
"privacy_passed": true,
"checksum_verified_before_document_refresh": true
},
"docs": {
"passed": true,
"speed_results": "placeholder requested"
}
},
"kernels": {
"passed": true,
"complete_runs": 3,
"runs": [
{
"exit_code": 0,
"counts": {
"passed": 9784,
"failed": 0,
"unsupported": 1793
},
"families": {
"MUL_MAT": {
"passed": 3918,
"failed": 0,
"unsupported": 420
},
"MUL_MAT_ID": {
"passed": 962,
"failed": 0,
"unsupported": 14
},
"FLASH_ATTN_EXT": {
"passed": 4004,
"failed": 0,
"unsupported": 1339
},
"SSM_CONV": {
"passed": 135,
"failed": 0,
"unsupported": 0
},
"GATED_DELTA_NET": {
"passed": 48,
"failed": 0,
"unsupported": 0
},
"RMS_NORM": {
"passed": 53,
"failed": 0,
"unsupported": 0
},
"ROPE": {
"passed": 450,
"failed": 0,
"unsupported": 20
},
"SOFT_MAX": {
"passed": 214,
"failed": 0,
"unsupported": 0
}
},
"failed_cases": [],
"passed": true,
"attempt": 4,
"stdout_sha256": "8c6fa2b880b8f196626891815d0e56387134c29bef786c9896388e46eddc785f",
"stderr_sha256": "f7b016d89dc0f27c6384e483758987c04fc856b6eff33cedd9ef183adf2cdb75",
"elapsed_s": 746.3958051204681,
"q5_1_cases": 59,
"q5_1_mul_mat_cases": 56
},
{
"exit_code": 0,
"counts": {
"passed": 9784,
"failed": 0,
"unsupported": 1793
},
"families": {
"MUL_MAT": {
"passed": 3918,
"failed": 0,
"unsupported": 420
},
"MUL_MAT_ID": {
"passed": 962,
"failed": 0,
"unsupported": 14
},
"FLASH_ATTN_EXT": {
"passed": 4004,
"failed": 0,
"unsupported": 1339
},
"SSM_CONV": {
"passed": 135,
"failed": 0,
"unsupported": 0
},
"GATED_DELTA_NET": {
"passed": 48,
"failed": 0,
"unsupported": 0
},
"RMS_NORM": {
"passed": 53,
"failed": 0,
"unsupported": 0
},
"ROPE": {
"passed": 450,
"failed": 0,
"unsupported": 20
},
"SOFT_MAX": {
"passed": 214,
"failed": 0,
"unsupported": 0
}
},
"failed_cases": [],
"passed": true,
"attempt": 5,
"stdout_sha256": "8c6fa2b880b8f196626891815d0e56387134c29bef786c9896388e46eddc785f",
"stderr_sha256": "f7b016d89dc0f27c6384e483758987c04fc856b6eff33cedd9ef183adf2cdb75",
"elapsed_s": 704.5498714447021,
"q5_1_cases": 59,
"q5_1_mul_mat_cases": 56
},
{
"exit_code": 0,
"counts": {
"passed": 9784,
"failed": 0,
"unsupported": 1793
},
"families": {
"MUL_MAT": {
"passed": 3918,
"failed": 0,
"unsupported": 420
},
"MUL_MAT_ID": {
"passed": 962,
"failed": 0,
"unsupported": 14
},
"FLASH_ATTN_EXT": {
"passed": 4004,
"failed": 0,
"unsupported": 1339
},
"SSM_CONV": {
"passed": 135,
"failed": 0,
"unsupported": 0
},
"GATED_DELTA_NET": {
"passed": 48,
"failed": 0,
"unsupported": 0
},
"RMS_NORM": {
"passed": 53,
"failed": 0,
"unsupported": 0
},
"ROPE": {
"passed": 450,
"failed": 0,
"unsupported": 20
},
"SOFT_MAX": {
"passed": 214,
"failed": 0,
"unsupported": 0
}
},
"failed_cases": [],
"passed": true,
"attempt": 7,
"stdout_sha256": "8c6fa2b880b8f196626891815d0e56387134c29bef786c9896388e46eddc785f",
"stderr_sha256": "f7b016d89dc0f27c6384e483758987c04fc856b6eff33cedd9ef183adf2cdb75",
"elapsed_s": 769.3762726783752,
"q5_1_cases": 59,
"q5_1_mul_mat_cases": 56
}
],
"counts": {
"passed": 29352,
"failed": 0,
"unsupported": 5379
},
"families": {
"MUL_MAT": {
"passed": 11754,
"failed": 0,
"unsupported": 1260
},
"MUL_MAT_ID": {
"passed": 2886,
"failed": 0,
"unsupported": 42
},
"FLASH_ATTN_EXT": {
"passed": 12012,
"failed": 0,
"unsupported": 4017
},
"SSM_CONV": {
"passed": 405,
"failed": 0,
"unsupported": 0
},
"GATED_DELTA_NET": {
"passed": 144,
"failed": 0,
"unsupported": 0
},
"RMS_NORM": {
"passed": 159,
"failed": 0,
"unsupported": 0
},
"ROPE": {
"passed": 1350,
"failed": 0,
"unsupported": 60
},
"SOFT_MAX": {
"passed": 642,
"failed": 0,
"unsupported": 0
}
},
"q5_1_cases": 177,
"excluded_attempts": {
"1": "owner interruption",
"2": "owner interruption",
"3": "owner interruption",
"6": "lease deferral; test never launched"
},
"q5_1_mul_mat_cases": 168,
"classification_rule": "MUL_MAT_ID and MUL_MAT_ID_* classified before MUL_MAT*; same corrected rule as initial comparison."
},
"model_agreement": {
"passed": true,
"scored_positions": 1020,
"context": 512,
"chunks": 4,
"mean_kld": 0.000597,
"p99_kld": 0.0162,
"maximum_kld": 0.040878,
"same_top_percent": 99.51,
"cpu_ppl": 1.155534,
"gpu_ppl": 1.154694,
"ppl_ratio": 0.999273,
"ppl_delta": -0.00084,
"corpus_sha256": "68c3808a113f74773f6c3c278173171358f2402d4e38e116ae6a03aca0263055",
"reference_log_probs_sha256": "4e71a366b8cc432cdf6b9c8c2add9cd8f69cd28b4fc6652ee72eab61acbfa223",
"same_top_count": 1015,
"thresholds": {
"mean_kld_max": 0.001,
"same_top_percent_min": 99,
"absolute_ppl_change_percent_max": 0.5
},
"base_format_note": "Upstream --kl-divergence-base stores scaled uint16 log probabilities; reported KLD is the upstream estimate, including its p > exp(-16) cutoff.",
"cpu_only_verified": true,
"gpu_offload_layers": [
41,
41
],
"flash_attention": "enabled",
"ppl_relative_change_percent": -0.07269366370872277,
"model": {
"filename": "MiniCPM-V-4_7-35B-A3B-Q4_K_M.gguf",
"sha256": "9ecf91e814d7e9adc64325b1b65f8628a9b1bb8da4f61a92869ca7489bfd888e",
"size_bytes": 21166255008
}
},
"historical_q5_1_failure": {
"op": "MUL_MAT",
"type_a": "q5_1",
"shape": {
"m": 16,
"n": 1,
"k": 256
},
"nmse": 0.00074011,
"threshold": 0.0005,
"threshold_multiple": 1.48022,
"input_captured": false,
"exact_cpu_replay_available": false,
"conditioning_proven": false,
"initial_supported_mul_mat_cases": 3918,
"initial_plus_fresh_mul_mat_cases": 15672,
"initial_plus_fresh_failures": 1,
"initial_plus_fresh_observed_failure_fraction": 6.380806533945891e-05,
"blocked_requirement": "Every failure must retain its input and be replayed on CPU with evidence establishing input conditioning."
},
"additional_reported_evidence": {
"independently_rerun_in_this_validation": false,
"v100_q5_1": {
"runs": 20,
"passed": 1120,
"failed": 0
},
"sm86_q5_1": {
"runs": 5,
"passed": 280,
"failed": 0
},
"not_in_fresh_gate_counts": true
},
"greedy": {
"is_gate": false,
"first_divergence_1_based": {
"reasoning": 26,
"prose": 12,
"code": null
},
"tokens_per_prompt": 64
},
"instrumentation": {
"original_test_source_sha256": "9acaf591366b892274dc5167308c2d9f6d82f257d1e9b9d9225be79d3e944b05",
"diagnostic_source_sha256": "3bbe160c550ee73193ec1d945c3154b9f10125c01f9de82c269a9c7c09f73a40",
"diagnostic_helper_sha256": "467e6d876f89fa203b6c185d81428bbb9470459c8193cd3306f5c52aa6ee1950",
"diagnostic_binary_sha256": "ff53ce0b103513584a3345a7a248fa9634c157d0a04fe58ad81d1d1385bbd7cf",
"production_unchanged": true,
"instrumentation": "Header include, CPU-only replay CLI branch, and capture calls immediately before existing comparison failure prints only",
"limitations": "Allocation/backend execution failures before comparison have no pair of outputs to capture. Fused intermediate outputs may not be materialized; the compared root output is valid."
},
"relocation": {
"files": [
{
"file": "libggml-base.so.0.27.0",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libggml-cpu.so.0.27.0",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libggml-cuda.so.0.27.0",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libggml.so.0.27.0",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libllama-bench-impl.so",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libllama-cli-impl.so",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libllama-common.so.0.7.0",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libllama-perplexity-impl.so",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libllama-quantize-impl.so",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libllama-server-impl.so",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libllama.so.0.7.0",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "libmtmd.so.0.7.0",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "llama-bench",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "llama-cli",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "llama-perplexity",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "llama-quantize",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "llama-server",
"build_matches": true,
"unpacked_matches": true
},
{
"file": "test-backend-ops",
"build_matches": true,
"unpacked_matches": true
}
],
"tools": [
{
"tool": "llama-server",
"probe": "--version",
"exit_code": 0,
"all_dependencies_found": true,
"no_build_dependencies": true
},
{
"tool": "llama-cli",
"probe": "--version",
"exit_code": 0,
"all_dependencies_found": true,
"no_build_dependencies": true
},
{
"tool": "llama-bench",
"probe": "--version",
"exit_code": 0,
"all_dependencies_found": true,
"no_build_dependencies": true
},
{
"tool": "llama-perplexity",
"probe": "--version",
"exit_code": 0,
"all_dependencies_found": true,
"no_build_dependencies": true
},
{
"tool": "llama-quantize",
"probe": "--help",
"exit_code": 1,
"all_dependencies_found": true,
"no_build_dependencies": true
},
{
"tool": "test-backend-ops",
"probe": "--help",
"exit_code": 1,
"all_dependencies_found": true,
"no_build_dependencies": true
}
],
"environment": {
"LD_LIBRARY_PATH": "unset",
"CUDA_VISIBLE_DEVICES": "empty",
"PATH": "/usr/bin:/bin"
},
"unpacked_folder_preexisting": true,
"runtime_only_limitation": "Host has a CUDA toolkit installed; probes and ELF dependencies use only runtime libraries, without nvcc in PATH. This is not an isolated runtime-only OS image.",
"passed": true,
"supported_version_probes_passed": 4,
"usage_probes_passed": 2
},
"binary_sha256": {
"libggml-base.so.0.27.0": "7dec032a7496b1ade3f9db26585c3eff5d7fe0940c694179e43d087edc5aa641",
"libggml-cpu.so.0.27.0": "9328d2600994cc61b76226b1865afc3f2e71945b7f2793e473fc84796df6ff36",
"libggml-cuda.so.0.27.0": "97acd2d0b82674e731605705a7d4c826604840bbcfc4eea7987286cdc9bc5a79",
"libggml.so.0.27.0": "9f39ebbcae29ee028ab45fec86b0f312dc6e3a0ec4db419c1deb86347c18d621",
"libllama-bench-impl.so": "084d5060baae43d8fae2a3de35d743183f124eace7e264d2657ea6c5547e19c0",
"libllama-cli-impl.so": "9246772a78adc8581adbed42b16bdc17ddc891479647dea73cad5906bfc16f52",
"libllama-common.so.0.7.0": "fd83ff307e92e64b32aca8799dde1eb4791ccc274e84ca8ae19e6d299533a199",
"libllama-perplexity-impl.so": "43170cc0162c08eb09054905476523cdd8292f6b2de959e733ccf01cbe5dce7e",
"libllama-quantize-impl.so": "55f2bf9cc8088be6078e90f3c826eac11a1cc89fc5e34821944e9b3f02a1c394",
"libllama-server-impl.so": "6b5fbb17ee399774b421c8190c87051e95a54177a5d9370edabd60e9094f43da",
"libllama.so.0.7.0": "f02e3993196df4f2c7403d53b3ce3afd6705678a941497c7250ff2d6c316210b",
"libmtmd.so.0.7.0": "997927073ce7c6601db58cdc57bfba5891ec6dca63d5a6374a8f87de51b672e1",
"llama-bench": "c542ade6b00f3e108efa8312776b17dbca8cceef45590845236328f67fb9ad7c",
"llama-cli": "00272bbaf8704619739c2ecbe970eb53a9d2c7544c26c8f41730e0dde0fa4a65",
"llama-perplexity": "b42d4afd0013c8d54c80cd84057beea784d094d9710f6fa3e9209fcce1862690",
"llama-quantize": "dded52ad19ad77f40fe678ad7b06627fc442cd437e3cd44b8c9f4c34b04115a7",
"llama-server": "0015eb3199723cfae5a15ec89bdbccc9fde8bd652c589179d134587804964b98",
"test-backend-ops": "bca163572bc5206896c68e2cab98612c97e31445a642aebdfbef8dd01d1685b7"
},
"initial_validation_sha256": "49d0be0dd26591fa042c54e07c62b642fbd37dedfeb358001238f58486dd465b",
"publication": "none",
"license": "MIT; preserve LICENSE and THIRD-PARTY-NOTICES"
}