Publish 3e536ca project-local runtime and inherited training config
Browse files
INSTALL.md
CHANGED
|
@@ -212,6 +212,12 @@ To reuse the current training environment, invoke the versioned
|
|
| 212 |
data, and delegates cache reuse to the existing `copy_gym_venvs.sh`, then checks
|
| 213 |
imports. It preserves the launcher's environment layout and starts no training.
|
| 214 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 215 |
For the pinned public dataset revision
|
| 216 |
`6f274cd4cbc27e656d7ddae7c3b591669866dc84`,
|
| 217 |
[`packaging/prepare_dataset.py`](packaging/prepare_dataset.py) combines extraction,
|
|
|
|
| 212 |
data, and delegates cache reuse to the existing `copy_gym_venvs.sh`, then checks
|
| 213 |
imports. It preserves the launcher's environment layout and starts no training.
|
| 214 |
|
| 215 |
+
When an execution scope excludes the existing system cache, use
|
| 216 |
+
`packaging/prepare_project_runtime.sh` to copy the required cached environments
|
| 217 |
+
into a fresh project directory. It generates an inherited trainer YAML that
|
| 218 |
+
changes only `env.nemo_gym.uv_venv_dir`, and verifies existing source/data and
|
| 219 |
+
frontend imports without editing the base trainer configuration or system cache.
|
| 220 |
+
|
| 221 |
For the pinned public dataset revision
|
| 222 |
`6f274cd4cbc27e656d7ddae7c3b591669866dc84`,
|
| 223 |
[`packaging/prepare_dataset.py`](packaging/prepare_dataset.py) combines extraction,
|
SOURCE_RELEASE.json
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
{
|
| 2 |
-
"archive": "spreadsheet-gym-integration-
|
| 3 |
-
"sha256": "
|
| 4 |
-
"bytes":
|
| 5 |
-
"source_commit": "
|
| 6 |
-
"files":
|
| 7 |
"branch": "tungks/feat/tool/spreadsheet-rl",
|
| 8 |
"reproducible_rebuild": true,
|
| 9 |
"archive_manifest_verified": true,
|
| 10 |
-
"validation": "
|
| 11 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"archive": "spreadsheet-gym-integration-3e536cae8dfa.tar.gz",
|
| 3 |
+
"sha256": "d1a37c134043ae03eba68a6bb4a7bb0f7e6aecadfe563ce90338e5b9f5298bed",
|
| 4 |
+
"bytes": 94271,
|
| 5 |
+
"source_commit": "3e536cae8dfa74805da4dfd33a9874e570e7fd80",
|
| 6 |
+
"files": 54,
|
| 7 |
"branch": "tungks/feat/tool/spreadsheet-rl",
|
| 8 |
"reproducible_rebuild": true,
|
| 9 |
"archive_manifest_verified": true,
|
| 10 |
+
"validation": "86 passed, 4 Linux/XLSX tests skipped on Mac; project-local runtime deployment pending"
|
| 11 |
}
|
SOURCE_TRANSPORT.json
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
{
|
| 2 |
"schema_version": 1,
|
| 3 |
-
"source_commit": "
|
| 4 |
"encoding": "base64",
|
| 5 |
-
"encoded_file": "spreadsheet-source-
|
| 6 |
-
"encoded_bytes":
|
| 7 |
-
"encoded_sha256": "
|
| 8 |
-
"archive": "spreadsheet-gym-integration-
|
| 9 |
-
"archive_bytes":
|
| 10 |
-
"archive_sha256": "
|
| 11 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"schema_version": 1,
|
| 3 |
+
"source_commit": "3e536cae8dfa74805da4dfd33a9874e570e7fd80",
|
| 4 |
"encoding": "base64",
|
| 5 |
+
"encoded_file": "spreadsheet-source-3e536cae8dfa.base64.txt",
|
| 6 |
+
"encoded_bytes": 125697,
|
| 7 |
+
"encoded_sha256": "20021049ea4850a3ff7de261423bee4598118d8d2587c11ebe5b2492f4e9f808",
|
| 8 |
+
"archive": "spreadsheet-gym-integration-3e536cae8dfa.tar.gz",
|
| 9 |
+
"archive_bytes": 94271,
|
| 10 |
+
"archive_sha256": "d1a37c134043ae03eba68a6bb4a7bb0f7e6aecadfe563ce90338e5b9f5298bed"
|
| 11 |
}
|
TRAINING_SETUP.md
CHANGED
|
@@ -15,6 +15,7 @@ paths and private keys stay outside public release artifacts.
|
|
| 15 |
| `start_service.py` | Existing runtime bundle, qualified catalog, node-local service root and sandbox root receipt; optional `--export-private-config` | Start/reuse the API; export its Gym override to shared storage with mode 0600 |
|
| 16 |
| `prepare_training_node.sh` | `TRAINER_ROOT NLTK_DATA TRAINER_CONFIG` followed by the Python options below | Use the existing activation/cache helper, install source/config, convert data and check existing frontend venv imports |
|
| 17 |
| `prepare_cached_frontends.py` | `--venv-root`, `--expected-hostname`, `--receipt` | Copy the existing `math_with_judge` and `simple_agent` caches for the two new frontends; no installs or venv aliases |
|
|
|
|
| 18 |
| `build_source_bundle.py` | Clean source checkout; `--output-dir` outside it | Small source archive and checksum, pinned to the exact Git commit |
|
| 19 |
|
| 20 |
Conversion, source installation and source packaging use only Python's standard
|
|
@@ -94,9 +95,32 @@ copies the existing `math_with_judge` cache for `spreadsheet`, and the existing
|
|
| 94 |
Complete existing destinations are reused; linked or incomplete destinations
|
| 95 |
are preserved and rejected. This creates real cache copies without installing
|
| 96 |
dependencies or modifying the project's helper. On a Bridge binding whose write
|
| 97 |
-
root excludes the cache directory,
|
| 98 |
-
|
| 99 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 100 |
|
| 101 |
Each invocation saves a private phase receipt and cache-helper log under a
|
| 102 |
unique setup directory. `TRAINING_NODE_PREPARED` means these setup checks passed;
|
|
|
|
| 15 |
| `start_service.py` | Existing runtime bundle, qualified catalog, node-local service root and sandbox root receipt; optional `--export-private-config` | Start/reuse the API; export its Gym override to shared storage with mode 0600 |
|
| 16 |
| `prepare_training_node.sh` | `TRAINER_ROOT NLTK_DATA TRAINER_CONFIG` followed by the Python options below | Use the existing activation/cache helper, install source/config, convert data and check existing frontend venv imports |
|
| 17 |
| `prepare_cached_frontends.py` | `--venv-root`, `--expected-hostname`, `--receipt` | Copy the existing `math_with_judge` and `simple_agent` caches for the two new frontends; no installs or venv aliases |
|
| 18 |
+
| `prepare_project_runtime.sh` | `TRAINER_ROOT NLTK_DATA`, then `--base-config`, `--runtime-root`, `--qualified-dir`, `--output-dir`, `--private-config`, `--expected-hostname` | Prepare project-local caches and an inherited YAML; verify existing integration/data without modifying their files |
|
| 19 |
| `build_source_bundle.py` | Clean source checkout; `--output-dir` outside it | Small source archive and checksum, pinned to the exact Git commit |
|
| 20 |
|
| 21 |
Conversion, source installation and source packaging use only Python's standard
|
|
|
|
| 95 |
Complete existing destinations are reused; linked or incomplete destinations
|
| 96 |
are preserved and rejected. This creates real cache copies without installing
|
| 97 |
dependencies or modifying the project's helper. On a Bridge binding whose write
|
| 98 |
+
root excludes the cache directory, use the project-local runtime below instead
|
| 99 |
+
of writing outside the binding policy. The standalone cache copier also accepts
|
| 100 |
+
`--source-root`, `--include-model`, and `--relocate-activation` for copying an
|
| 101 |
+
existing cache into a new location.
|
| 102 |
+
|
| 103 |
+
## Project-local runtime under an existing Bridge scope
|
| 104 |
+
|
| 105 |
+
`prepare_project_runtime.sh` sources the existing import activation and invokes
|
| 106 |
+
`prepare_project_runtime.py`. The active trainer must already have the matching
|
| 107 |
+
spreadsheet frontend sources, config paths, and verified converted data. It
|
| 108 |
+
checks these prerequisites without installing source, patching the base YAML,
|
| 109 |
+
recalculating workbooks, or recreating data.
|
| 110 |
+
|
| 111 |
+
All persistent writes are below the fresh `--runtime-root`. It copies existing
|
| 112 |
+
`math_with_judge`, `simple_agent`, and `vllm_model` caches from
|
| 113 |
+
`--source-cache-root` (default `/opt/gym_venvs`) into the project's resource,
|
| 114 |
+
agent, and model cache directories. Only copied Bash activation scripts are
|
| 115 |
+
retargeted. Import probes require the copied interpreter prefixes and matching
|
| 116 |
+
Ray/OpenAI versions. Python bytecode writes are disabled during preparation and
|
| 117 |
+
probes.
|
| 118 |
+
|
| 119 |
+
The new `spreadsheet.yaml` inherits `--base-config` and overrides only
|
| 120 |
+
`env.nemo_gym.uv_venv_dir`. Setup verifies the complete effective configuration
|
| 121 |
+
equals the base with that single change. Use this generated YAML when launching
|
| 122 |
+
training. `setup.json` pins the base/config/source/data hashes and live service
|
| 123 |
+
health. `TRAINING_NODE_PREPARED` confirms preparation; no training is launched.
|
| 124 |
|
| 125 |
Each invocation saves a private phase receipt and cache-helper log under a
|
| 126 |
unique setup directory. `TRAINING_NODE_PREPARED` means these setup checks passed;
|
spreadsheet-gym-integration-3e536cae8dfa.tar.gz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d1a37c134043ae03eba68a6bb4a7bb0f7e6aecadfe563ce90338e5b9f5298bed
|
| 3 |
+
size 94271
|
spreadsheet-gym-integration-3e536cae8dfa.tar.gz.sha256
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:31b1cf469f993525a593d5e22d6c891fcede70f631685ae2a4a8c49d5ea29788
|
| 3 |
+
size 114
|
spreadsheet-source-3e536cae8dfa.base64.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|