diff --git a/.github/ci/reference/proposed_model_ports/afm45b.json b/.github/ci/reference/proposed_model_ports/afm45b.json new file mode 100644 index 00000000..6012c210 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/afm45b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "arcee-ai/AFM-4.5B", + "revision": "e84d19711eca5817c1608cb344d931bf05c966dd", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "2de4692fc3404ba302701753f4110610d3d8f88cf49294d5ce0f176e01cc8871", + "quantization": "Q8_0", + "hf_repo": "arcee-ai/AFM-4.5B-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "arcee" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/afm45b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/apertus_8b.json b/.github/ci/reference/proposed_model_ports/apertus_8b.json new file mode 100644 index 00000000..032c8986 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/apertus_8b.json @@ -0,0 +1,44 @@ +{ + "schema_version": 1, + "model": { + "repo": "swiss-ai/Apertus-8B-Instruct-2509", + "revision": "b946d40447b2b597999b9c86d44bee0b452c919f", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "823a2ce1b89f8e31c197167e144fb4e5220f0adb44132b8e54ccdb2dd5bedae7", + "quantization": "Q8_0", + "hf_repo": "DevQuasar/swiss-ai.Apertus-8B-Instruct-2509-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { + "general_architecture": "apertus" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 7.8747, + "host_cpu_perplexity_error": 1.46638, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/apertus_8b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/chatglm3_6b.json b/.github/ci/reference/proposed_model_ports/chatglm3_6b.json new file mode 100644 index 00000000..a336b7b8 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/chatglm3_6b.json @@ -0,0 +1,44 @@ +{ + "schema_version": 1, + "model": { + "repo": "zai-org/chatglm3-6b", + "revision": "e9e0406d062cdb887444fe5bd546833920abd4ac", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "805e3761a997486eb08f45e0f74efa5c1cc0be8475afa2b170263b490a3ba8c8", + "quantization": "Q8_0", + "hf_repo": "hellork/chatglm3-6b-128k-Q8_0-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { + "general_architecture": "chatglm" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 15.91, + "host_cpu_perplexity_error": 3.3745, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/chatglm3_6b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/codeshell_7b.json b/.github/ci/reference/proposed_model_ports/codeshell_7b.json new file mode 100644 index 00000000..cc8a4a9b --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/codeshell_7b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "WisdomShell/CodeShell-7B-Chat", + "revision": "3cb06f589b7b1e2f8e728c77280b1114191d24de", + "license": "codeshell-license" + }, + "baseline_artifacts": { + "model": { + "sha256": "207e1ee777455b17d3e58f3174dead52f1c8d007690c96679a7deba355284876", + "quantization": "Q8_0", + "hf_repo": "DevQuasar/WisdomShell.Shell-7B-Chat-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "codeshell" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/codeshell_7b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/cohere2_7b.json b/.github/ci/reference/proposed_model_ports/cohere2_7b.json new file mode 100644 index 00000000..40cb5c13 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/cohere2_7b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "CohereLabs/c4ai-command-r7b-12-2024", + "revision": "4f3d0aa6856e322f2f4480fe65420d5d53d297b8", + "license": "cc-by-nc-4.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "cd7281ec7974bb5810d8f6c922801d6fff3e8bd761073942757dcef990ca9b4d", + "quantization": "Q8_0", + "hf_repo": "bartowski/c4ai-command-r7b-12-2024-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "cohere2" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/cohere2_7b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/decilm7b.json b/.github/ci/reference/proposed_model_ports/decilm7b.json new file mode 100644 index 00000000..9776451e --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/decilm7b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "Deci/DeciLM-7B-instruct", + "revision": "4adc7aa9efe61b47b0a98b2cc94527d9c45c3b4f", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "1b60beae4c9d5d7131f846f7545cb386e20d8f2c613844a0142ee701537dbfe3", + "quantization": "Q8_0", + "hf_repo": "Deci/DeciLM-7B-instruct-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "deci" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/decilm7b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/ernie45_03b.json b/.github/ci/reference/proposed_model_ports/ernie45_03b.json new file mode 100644 index 00000000..b753899b --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/ernie45_03b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "baidu/ERNIE-4.5-0.3B-PT", + "revision": "b565cf6caebdb7a1eadf00100857b1ed5e044f12", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "022ce1cfc46a2818a96b8a683aec0944047cac6745f35aa5a2a48729a58caa74", + "quantization": "Q8_0", + "hf_repo": "bartowski/baidu_ERNIE-4.5-0.3B-PT-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "ernie4_5" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/ernie45_03b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/falcon7b.json b/.github/ci/reference/proposed_model_ports/falcon7b.json new file mode 100644 index 00000000..239f9589 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/falcon7b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "tiiuae/falcon-7b", + "revision": "ec89142b67d748a1865ea4451372db8313ada0d8", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "47e0a8ef8eb3d55765f7d412a1d9b816159fece3e9f87ce79caa7e6d9b2341a8", + "quantization": "Q8_0", + "hf_repo": "maddes8cht/tiiuae-falcon-7b-gguf" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "falcon" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/falcon7b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/falcon_h1_1_5b.json b/.github/ci/reference/proposed_model_ports/falcon_h1_1_5b.json new file mode 100644 index 00000000..e274207a --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/falcon_h1_1_5b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "tiiuae/Falcon-H1-1.5B-Instruct", + "revision": "80ebc50d7799a440b96c93bb6686a3924a09b0cb", + "license": "falcon-llm-license" + }, + "baseline_artifacts": { + "model": { + "sha256": "0044322ad5aaea5ccc2264f83f430f4ab2d20acc6db71a12d79b3f5acb0f69d5", + "quantization": "Q8_0", + "hf_repo": "tiiuae/Falcon-H1-1.5B-Instruct-GGUF" + }, + "note": "Official quantization repo from the model's own authors." + }, + "architecture": { "general_architecture": "falcon-h1" }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 11.6708, + "host_cpu_perplexity_error": 2.60807, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/falcon_h1_1_5b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/gemma2_2b.json b/.github/ci/reference/proposed_model_ports/gemma2_2b.json new file mode 100644 index 00000000..2e7616f9 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/gemma2_2b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "google/gemma-2-2b-it", + "revision": "299a8560bedf22ed1c72a8a11e7dce4a7f9f51f8", + "license": "gemma" + }, + "baseline_artifacts": { + "model": { + "sha256": "2d448a9aab894b8e8e18168cf3f490cb9f65632222f29f93514ac9ecc754debe", + "quantization": "Q8_0", + "hf_repo": "bartowski/gemma-2-2b-it-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "gemma2" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/gemma2_2b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/glm4_9b.json b/.github/ci/reference/proposed_model_ports/glm4_9b.json new file mode 100644 index 00000000..3c96e56d --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/glm4_9b.json @@ -0,0 +1,44 @@ +{ + "schema_version": 1, + "model": { + "repo": "zai-org/GLM-4-9B-0414", + "revision": "645b8482494e31b6b752272bf7f7f273ef0f3caf", + "license": "mit" + }, + "baseline_artifacts": { + "model": { + "sha256": "7b4ea2795934ca05dc409251dddd289160a194a4d920b575d1de4516808bb50d", + "quantization": "Q8_0", + "hf_repo": "bartowski/THUDM_GLM-4-9B-0414-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { + "general_architecture": "glm4" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 12.2894, + "host_cpu_perplexity_error": 2.42663, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/glm4_9b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/gpt2.json b/.github/ci/reference/proposed_model_ports/gpt2.json new file mode 100644 index 00000000..f0f09a53 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/gpt2.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "openai-community/gpt2", + "revision": "607a30d783dfa663caf39e06633721c8d4cfcd7e", + "license": "mit" + }, + "baseline_artifacts": { + "model": { + "sha256": "4d1df054e273ac10b3a31a2f90660426212d033156bb0b530e7517662fbe0c32", + "quantization": "Q8_0", + "hf_repo": "igorbkz/gpt2-Q8_0-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "gpt2" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/gpt2/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/granite4_h_micro.json b/.github/ci/reference/proposed_model_ports/granite4_h_micro.json new file mode 100644 index 00000000..23283bea --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/granite4_h_micro.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "ibm-granite/granite-4.0-h-micro", + "revision": "d5f01a3ea75f088947be3aae039f4ad52837dfde", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "a009111abf2865b7aad1e66326a6c772cddc29bccd22898f470292068b27bb59", + "quantization": "Q8_0", + "hf_repo": "ibm-granite/granite-4.0-h-micro-GGUF" + }, + "note": "Official quantization repo from the model's own authors." + }, + "architecture": { "general_architecture": "granitehybrid" }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 13.2058, + "host_cpu_perplexity_error": 2.78765, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/granite4_h_micro/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/granite_3_1b_a400m.json b/.github/ci/reference/proposed_model_ports/granite_3_1b_a400m.json new file mode 100644 index 00000000..3118d4d5 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/granite_3_1b_a400m.json @@ -0,0 +1,44 @@ +{ + "schema_version": 1, + "model": { + "repo": "ibm-granite/granite-3.0-1b-a400m-instruct", + "revision": "ffec3c35bdfd97a06f0b4cd5fcc92cd9b1584445", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "8c37dd0c10b73e9304b98a242be4adcd3050c09e9042c4862d49e2cfccf35411", + "quantization": "Q8_0", + "hf_repo": "bartowski/granite-3.0-1b-a400m-instruct-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { + "general_architecture": "granitemoe" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 5.7635, + "host_cpu_perplexity_error": 0.84793, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/granite_3_1b_a400m/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/internlm2_5_18b.json b/.github/ci/reference/proposed_model_ports/internlm2_5_18b.json new file mode 100644 index 00000000..1444d6f8 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/internlm2_5_18b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "internlm/internlm2_5-1_8b-chat", + "revision": "57dcbc5455bf050d27a80f0ebd7808387caa714d", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "8526cc24717fcab32b20540c546f8c23a6ea3ff40b86f421a0cd060c8123e8b2", + "quantization": "Q8_0", + "hf_repo": "internlm/internlm2_5-1_8b-chat-gguf" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "internlm2" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/internlm2_5_18b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/maincoder_1b.json b/.github/ci/reference/proposed_model_ports/maincoder_1b.json new file mode 100644 index 00000000..1f33b580 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/maincoder_1b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "Maincode/Maincoder-1B", + "revision": "088ec98640bdeb105f46a9ef6a1370ed5d0d2ea5", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "43cd43d942de46327ff82a77e63717ef9a34d97787a3c5dcd309a9da3978a668", + "quantization": "Q8_0", + "hf_repo": "Maincode/Maincoder-1B-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "maincoder" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/maincoder_1b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/mamba_1_4b.json b/.github/ci/reference/proposed_model_ports/mamba_1_4b.json new file mode 100644 index 00000000..b9bcce23 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/mamba_1_4b.json @@ -0,0 +1,44 @@ +{ + "schema_version": 1, + "model": { + "repo": "state-spaces/mamba-1.4b-hf", + "revision": "6e46eae61c27280517feef46f536d16b91076f08", + "license": "unspecified" + }, + "baseline_artifacts": { + "model": { + "sha256": "a5dde81ce41c34213bf26bcaf2ba26cd1c6fde20fd20b5af6a4a76a5272478bc", + "quantization": "Q6_K", + "hf_repo": "RichardErkhov/state-spaces_-_mamba-1.4b-hf-gguf" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { + "general_architecture": "mamba" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 11.7985, + "host_cpu_perplexity_error": 1.97189, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/mamba_1_4b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/minicpm3_4b.json b/.github/ci/reference/proposed_model_ports/minicpm3_4b.json new file mode 100644 index 00000000..8e450f16 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/minicpm3_4b.json @@ -0,0 +1,44 @@ +{ + "schema_version": 1, + "model": { + "repo": "openbmb/MiniCPM3-4B", + "revision": "d6b14ddaefdb11c624dd75c3c779549bc90b08cb", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "a540961f131d6ab9616b6cb1c266be86735afe84907a445ad68e1e3e9422d122", + "quantization": "Q8_0", + "hf_repo": "QuantFactory/MiniCPM3-4B-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { + "general_architecture": "minicpm3" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 11.8752, + "host_cpu_perplexity_error": 2.29542, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/minicpm3_4b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/minicpm5_1b.json b/.github/ci/reference/proposed_model_ports/minicpm5_1b.json new file mode 100644 index 00000000..5fafb389 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/minicpm5_1b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "openbmb/MiniCPM5-1B", + "revision": "4e9de7a0778dc1c362e983e6858f0e77542cbdca", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "0dc7638539067268774c275a14a6ec9c7e01f7eeb2cff606c8590361fa527e4c", + "quantization": "Q8_0", + "hf_repo": "openbmb/MiniCPM5-1B-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "minicpm" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/minicpm5_1b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/nemotron_mini_4b.json b/.github/ci/reference/proposed_model_ports/nemotron_mini_4b.json new file mode 100644 index 00000000..254c6e37 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/nemotron_mini_4b.json @@ -0,0 +1,44 @@ +{ + "schema_version": 1, + "model": { + "repo": "nvidia/Nemotron-Mini-4B-Instruct", + "revision": "791833e92ebddb0bc2c1007f6d2b6764f886a2ae", + "license": "nvidia-community-model-license" + }, + "baseline_artifacts": { + "model": { + "sha256": "8ed6d148ded733d4401495e44dc834ffa79e8cfee6f44d8fb60808dddd78b8eb", + "quantization": "Q8_0", + "hf_repo": "bartowski/Nemotron-Mini-4B-Instruct-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { + "general_architecture": "nemotron" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 10.5717, + "host_cpu_perplexity_error": 2.16348, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/nemotron_mini_4b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/olmo_1b.json b/.github/ci/reference/proposed_model_ports/olmo_1b.json new file mode 100644 index 00000000..151ef094 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/olmo_1b.json @@ -0,0 +1,44 @@ +{ + "schema_version": 1, + "model": { + "repo": "allenai/OLMo-1B-hf", + "revision": "aee7752d9c08ee4775e9b0091426d8410e8f6a89", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "11ad66bb4b0c4b9d4b40ccef351a506e322fc8265b87e3638b18802768d8875e", + "quantization": "Q8_0", + "hf_repo": "RichardErkhov/allenai_-_OLMo-1B-hf-gguf" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { + "general_architecture": "olmo" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 11.5101, + "host_cpu_perplexity_error": 1.96062, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/olmo_1b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/openelm_1_1b.json b/.github/ci/reference/proposed_model_ports/openelm_1_1b.json new file mode 100644 index 00000000..df89d4cd --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/openelm_1_1b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "apple/OpenELM-1_1B", + "revision": "ee559a10b14895dde9f8cfde3fdc77b3ff0dbc0f", + "license": "apple-sample-code-license" + }, + "baseline_artifacts": { + "model": { + "sha256": "13dc6676b2355d0356d7892d8b34d8518233d79a28f9ea1d14f94e81d7decad5", + "quantization": "Q8_0", + "hf_repo": "LiteLLMs/OpenELM-1_1B-Instruct-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { "general_architecture": "openelm" }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 14.4998, + "host_cpu_perplexity_error": 3.00329, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/openelm_1_1b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/phi15.json b/.github/ci/reference/proposed_model_ports/phi15.json new file mode 100644 index 00000000..f98aa182 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/phi15.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "microsoft/phi-1_5", + "revision": "77aa61eeac94fbf33d492b9f2744c98b42d5b5eb", + "license": "microsoft-research-license" + }, + "baseline_artifacts": { + "model": { + "sha256": "e8c26615319e1141348b8534641da54d58e02b7baa01ee611b9c69cc07bf43fd", + "quantization": "Q8_0", + "hf_repo": "TKDKid1000/phi-1_5-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "phi2" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/phi15/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/phi3_mini.json b/.github/ci/reference/proposed_model_ports/phi3_mini.json new file mode 100644 index 00000000..382c3891 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/phi3_mini.json @@ -0,0 +1,44 @@ +{ + "schema_version": 1, + "model": { + "repo": "microsoft/Phi-3-mini-4k-instruct", + "revision": "f39ac1d28e925b323eae81227eaba4464caced4e", + "license": "mit" + }, + "baseline_artifacts": { + "model": { + "sha256": "8a83c7fb9049a9b2e92266fa7ad04933bb53aa1e85136b7b30f1b8000ff2edef", + "quantization": "Q4", + "hf_repo": "microsoft/Phi-3-mini-4k-instruct-gguf" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { + "general_architecture": "phi3" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 9.485, + "host_cpu_perplexity_error": 1.86142, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/phi3_mini/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/plamo3_nict_8b.json b/.github/ci/reference/proposed_model_ports/plamo3_nict_8b.json new file mode 100644 index 00000000..481fca27 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/plamo3_nict_8b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "pfnet/plamo-3-nict-8b-base", + "revision": "7adeb6e9fc79016d4e99e2e3cb8fde6441b42366", + "license": "plamo-community-license" + }, + "baseline_artifacts": { + "model": { + "sha256": "339e0c93206fbf495dd35d47ec2d942286c26a0cb70294546a75588eb5129bbb", + "quantization": "Q8_0", + "hf_repo": "mmnga-o/plamo-3-nict-8b-base-gguf" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "plamo3" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/plamo3_nict_8b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/refact16b.json b/.github/ci/reference/proposed_model_ports/refact16b.json new file mode 100644 index 00000000..91465430 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/refact16b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "smallcloudai/Refact-1_6B-fim", + "revision": "a2b605b6d76ce09732ef6f54265b3c0a8393c08e", + "license": "bigscience-openrail-m" + }, + "baseline_artifacts": { + "model": { + "sha256": "2298c7e2ad48d17e9db5c46e783b55dfe9e8ed99a00364b9c1f3bd14c0f6519f", + "quantization": "Q8_0", + "hf_repo": "oblivious/Refact-1.6B-fim-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "refact" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/refact16b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/smollm3_3b.json b/.github/ci/reference/proposed_model_ports/smollm3_3b.json new file mode 100644 index 00000000..0d5e7993 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/smollm3_3b.json @@ -0,0 +1,44 @@ +{ + "schema_version": 1, + "model": { + "repo": "HuggingFaceTB/SmolLM3-3B", + "revision": "a07cc9a04f16550a088caea529712d1d335b0ac1", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "8aa8cc74656137174a1988d993b00828e65a86fd68773412b632a75aa1373248", + "quantization": "Q8_0", + "hf_repo": "ggml-org/SmolLM3-3B-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used." + }, + "architecture": { + "general_architecture": "smollm3" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "measured_host_cpu_only", + "host_cpu_perplexity": 12.3653, + "host_cpu_perplexity_error": 2.40724, + "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "measured_host_cpu_only" + }, + "recipe": "ported_models/smollm3_3b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/stablelm2_16b.json b/.github/ci/reference/proposed_model_ports/stablelm2_16b.json new file mode 100644 index 00000000..da4ab672 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/stablelm2_16b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "stabilityai/stablelm-2-1_6b", + "revision": "f499ead74c53749bd93cebc6ce8bc0d7bdf1eaef", + "license": "stabilityai-stablelm-2-1_6b" + }, + "baseline_artifacts": { + "model": { + "sha256": "fcdeda8182b99b007cb2c69b216e31efd21b78a4588188b076fa92821f3fee20", + "quantization": "Q8_0", + "hf_repo": "afrideva/stablelm-2-1_6b-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "stablelm" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/stablelm2_16b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/starcoder2_3b.json b/.github/ci/reference/proposed_model_ports/starcoder2_3b.json new file mode 100644 index 00000000..c7a8cbb2 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/starcoder2_3b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "bigcode/starcoder2-3b", + "revision": "733247c55e3f73af49ce8e9c7949bf14af205928", + "license": "bigcode-openrail-m" + }, + "baseline_artifacts": { + "model": { + "sha256": "6ed1dfa70e4695fbbc1cb781c4baa93da347ed25516395d48eaf99e9e86c4bc8", + "quantization": "Q8_0", + "hf_repo": "QuantFactory/starcoder2-3b-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "starcoder2" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/starcoder2_3b/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/tiny_starcoder_py.json b/.github/ci/reference/proposed_model_ports/tiny_starcoder_py.json new file mode 100644 index 00000000..f51a4e42 --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/tiny_starcoder_py.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "bigcode/tiny_starcoder_py", + "revision": "8547527bef0bc927268c1653cce6948c5c242dd1", + "license": "bigcode-openrail-m" + }, + "baseline_artifacts": { + "model": { + "sha256": "ffa756aaa62050ffeea4f5b8897c54aa1a4d9c9ee0783e6ccc07f10eff8b077a", + "quantization": "Q8_0", + "hf_repo": "RichardErkhov/bigcode_-_tiny_starcoder_py-gguf" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "starcoder" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/tiny_starcoder_py/docs/RECIPE.md" +} diff --git a/.github/ci/reference/proposed_model_ports/xverse7b.json b/.github/ci/reference/proposed_model_ports/xverse7b.json new file mode 100644 index 00000000..564c598c --- /dev/null +++ b/.github/ci/reference/proposed_model_ports/xverse7b.json @@ -0,0 +1,42 @@ +{ + "schema_version": 1, + "model": { + "repo": "xverse/XVERSE-7B-Chat", + "revision": "60acc8c453c067b54df88be98bfdf60585ab5441", + "license": "apache-2.0" + }, + "baseline_artifacts": { + "model": { + "sha256": "1f2cdfeedcf3e2fe37ba3436d2a879b3f5f5a71b9ebc2ba2e888f8a3e046388d", + "quantization": "Q8_0", + "hf_repo": "xverse/XVERSE-7B-Chat-GGUF" + }, + "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one." + }, + "architecture": { + "general_architecture": "xverse" + }, + "runtime": { + "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et", + "required_device": "ET", + "required_gpu_layers": 99, + "require_full_offload": true + }, + "quality": { + "perplexity": { + "corpus_artifact": "wikitext2_raw_test", + "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08", + "context_size": 128, + "batch_size": 128, + "ubatch_size": 128, + "chunks": 4, + "measurement_status": "pending_host_or_board_run", + "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed." + } + }, + "correctness": { + "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness", + "measurement_status": "pending_host_or_board_run" + }, + "recipe": "ported_models/xverse7b/docs/RECIPE.md" +}