starforge-cli 0.1.6__py3-none-any.whl

starforge_cli/spec_builder.py

@@ -0,0 +1,372 @@
+"""从 CLI 参数构建 JobSpec。
+
+客户端构建完整 JobSpec，并在**本地**按 recipe 声明预校验 —— 超参拼错、
+取值越界在敲回车那一刻就能看到，不必等上传完、排完队、跑起来才发现。
+
+本地校验用的是与服务端**同一份** recipe 目录（都来自 starforge-core），所以不会
+出现「本地说没问题、服务端说不行」。
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any, Optional
+
+from starforge_core import __version__ as CORE_VERSION
+from starforge_core.contract import (
+    ArtifactRef,
+    DataRef,
+    DataSpec,
+    FrameworkRef,
+    JobSpec,
+    JobSpecBody,
+    LifecycleSpec,
+    Metadata,
+    ModelSpec,
+    PluginUse,
+    Provenance,
+    RecipeRef,
+    ResourcePool,
+    ResourceSpec,
+    SourceSpec,
+    SpecError,
+    exp_basename,
+)
+from starforge_core.recipes import RuntimeArtifact, get_recipe
+
+
+def parse_set(items: list[str]) -> dict[str, Any]:
+    """把 `--set k=v` 解析成超参字典，并按字面量推断类型。
+
+    类型推断是必要的：recipe schema 会拒绝 "500"（字符串）作为 max_num_steps，
+    而命令行给过来的一切都是字符串。推断规则刻意保守 —— 只认 bool/int/float，
+    其余保持字符串，避免把 "1e-5" 之外的东西误判。
+    """
+    out: dict[str, Any] = {}
+    for raw in items or []:
+        if "=" not in raw:
+            raise SpecError(f"--set 需要 key=value 形式，收到 {raw!r}")
+        key, _, val = raw.partition("=")
+        key, val = key.strip(), val.strip()
+        if not key:
+            raise SpecError(f"--set 的 key 不能为空: {raw!r}")
+        out[key] = _literal(val)
+    return out
+
+
+def _literal(v: str) -> Any:
+    low = v.lower()
+    if low in ("true", "yes", "on"):
+        return True
+    if low in ("false", "no", "off"):
+        return False
+    try:
+        return int(v)
+    except ValueError:
+        pass
+    try:
+        return float(v)
+    except ValueError:
+        return v
+
+
+def parse_pool(raw: str) -> ResourcePool:
+    """解析内部资源池字符串 `name:series:nodes:gpus_per_node`。
+
+    用户不再手写这个格式——CLI 把 `--profile 名称[:总卡数]` 经注册表物化成
+    这种字符串后传进来（见 materialize_pools）；保留字符串形态是为了
+    build_spec 的入参与测试稳定。
+    """
+    parts = [p.strip() for p in raw.split(":")]
+    if len(parts) != 4:
+        raise SpecError(
+            f"资源池需要 name:series:nodes:gpus_per_node 四段，收到 {raw!r}"
+            "（例：train:h200:1:4）"
+        )
+    name, series, nodes, per_node = parts
+    for label, v in (("nodes", nodes), ("gpus_per_node", per_node)):
+        if not v.isdigit() or int(v) <= 0:
+            raise SpecError(f"资源池的 {label} 应为正整数，收到 {v!r}")
+    return ResourcePool(name=name, series=series, nodes=int(nodes), gpus_per_node=int(per_node))
+
+
+def parse_profile_expr(raw: str) -> tuple[str, str, Optional[int]]:
+    """解析统一资源参数 `[role=]名称[:总卡数]` → (role, profile_name, gpus)。
+
+    例：h200 → ("", "h200", None)；h200:4 → ("", "h200", 4)；
+        rollout=h100:2 → ("rollout", "h100", 2)。
+    role 前缀是异构多池的扩展位：不同角色落在不同卡型/池上。
+    """
+    s = (raw or "").strip()
+    role, _, rest = s.partition("=") if "=" in s else ("", "", s)
+    role = role.strip()
+    name, sep, gpus_s = rest.partition(":")
+    name, gpus_s = name.strip(), gpus_s.strip()
+    if not name:
+        raise SpecError(f"--profile 需要 [role=]名称[:总卡数] 形式，收到 {raw!r}")
+    if not sep:
+        return role, name, None
+    if not gpus_s.isdigit() or int(gpus_s) <= 0:
+        raise SpecError(f"--profile 的总卡数应为正整数，收到 {raw!r}（例：h200:4）")
+    return role, name, int(gpus_s)
+
+
+def _shape_for(name: str, gpus: Optional[int], entry: dict) -> tuple[int, int]:
+    """由 profile 默认形状与可选的总卡数覆盖推导 (nodes, gpus_per_node)。
+
+    不超过单节点容量 → 单节点 N 卡；超过则必须是整节点的倍数——部分占用
+    多个节点的形状没有对应的调度语义，宁可报错也不猜。
+    """
+    per_node = int(entry.get("gpus_per_node") or 1)
+    if gpus is None:
+        return int(entry.get("num_nodes") or 1), per_node
+    if gpus <= per_node:
+        return 1, gpus
+    if gpus % per_node == 0:
+        return gpus // per_node, per_node
+    raise SpecError(
+        f"profile「{name}」每节点 {per_node} 卡：要 {gpus} 卡须不超过 {per_node}"
+        f"（单节点）或为 {per_node} 的整数倍（整节点）"
+    )
+
+
+def materialize_pools(
+    exprs: list[str], registry: dict[str, dict]
+) -> tuple[str, list[str], list[str]]:
+    """把 `--profile [role=]名称[:总卡数]` 物化成 (作业 profile, pools, roles)。
+
+    series 与默认形状一律取服务端注册表，用户没有途径写出「钉在 A 卡、
+    账记 B 卡」的 spec。单条目 → 单池 "all"（角色由 build_spec 自动推断）；
+    多条目是异构多池扩展位：每条必须带 role= 前缀，池名即角色名。
+    """
+    if not exprs:
+        raise SpecError("至少需要一个 --profile（`sf status` 查看可用值）")
+    parsed = [parse_profile_expr(e) for e in exprs]
+
+    known = "、".join(sorted(registry)) or "（注册表为空）"
+    for _, name, _ in parsed:
+        if name not in registry:
+            raise SpecError(f"未知的 profile「{name}」；可用：{known}")
+
+    if len(parsed) == 1:
+        role, name, gpus = parsed[0]
+        entry = registry[name]
+        nodes, per_node = _shape_for(name, gpus, entry)
+        pool_name = role or "all"
+        pools = [f"{pool_name}:{entry.get('series') or name}:{nodes}:{per_node}"]
+        roles = [f"{role}:{pool_name}"] if role else []
+        return name, pools, roles
+
+    missing = [raw for raw, (role, _, _) in zip(exprs, parsed) if not role]
+    if missing:
+        raise SpecError(
+            f"多个 --profile 时每条都要 role= 前缀（角色到池的映射），缺前缀：{missing[0]!r}"
+            "（例：--profile train=h200:8 --profile rollout=h100:2）"
+        )
+    seen: set[str] = set()
+    pools, roles = [], []
+    for role, name, gpus in parsed:
+        if role in seen:
+            raise SpecError(f"角色 {role!r} 重复声明")
+        seen.add(role)
+        entry = registry[name]
+        nodes, per_node = _shape_for(name, gpus, entry)
+        pools.append(f"{role}:{entry.get('series') or name}:{nodes}:{per_node}")
+        roles.append(f"{role}:{role}")
+    # 作业级 env/overrides/pin 目前是单 profile 语义，取第一条的 profile。
+    return parsed[0][1], pools, roles
+
+
+def build_spec(
+    exp_rel: str,
+    *,
+    recipe: str,
+    framework_version: str = "",
+    user: str = "",
+    project: str = "",
+    display_name: str = "",
+    sets: Optional[list[str]] = None,
+    pools: Optional[list[str]] = None,
+    roles: Optional[list[str]] = None,
+    init_from: str = "",
+    on_success: Optional[list[str]] = None,
+    image: str = "",
+    observability_url: str = "",
+    base_model: str = "",
+    train_data: str = "",
+    validation_data: str = "",
+    train_dataset: str = "",
+    validation_dataset: str = "",
+    operation: str = "train",
+    provenance: Optional[dict] = None,
+    validate: bool = True,
+    plugin_uses: Optional[list[dict]] = None,
+) -> JobSpec:
+    """构建并（默认）本地校验 JobSpec。
+
+    校验失败抛 SpecError —— CLI 捕获后打印可读错误，不发起任何网络请求。
+    """
+    # get_recipe 的报错已列出可用值，并对无前缀裸名给出两段式提示，直接透传。
+    r = get_recipe(recipe)
+    selected_framework_version = framework_version.strip() or r.runtime.default_version
+    selected_runtime = r.runtime.resolve(selected_framework_version)
+    requested_image = image.strip()
+    if requested_image:
+        RuntimeArtifact.from_dict(
+            {"kind": "oci", "reference": requested_image},
+            field_path="spec.framework.image",
+        )
+    if r.framework == "custom":
+        selected_image = requested_image
+    else:
+        # 空 = 服务端用 catalog / runtime registry；有值 = 按作业覆盖（服务端再查 allowlist）
+        selected_image = requested_image
+
+    observability = r.adapter_options.get("observability")
+    if observability == "external" and not observability_url.strip():
+        raise SpecError(
+            f"方法 {r.id} 使用 external observability，必须显式提供 --observability-url"
+        )
+    if observability == "platform" and observability_url.strip():
+        raise SpecError(f"方法 {r.id} 使用 platform observability，不允许 --observability-url")
+    if r.framework == "custom" and not selected_image:
+        raise SpecError("custom recipe 要求显式提供 --image")
+    if r.framework in {"verl", "trl"} and operation == "train":
+        missing = [
+            flag
+            for flag, value in (
+                ("--model", base_model),
+                ("--train-data", train_data),
+                ("--validation-data", validation_data),
+            )
+            if not value.strip()
+        ]
+        if missing:
+            raise SpecError(f"{r.framework} recipe 要求显式提供：{' '.join(missing)}")
+
+    # 平台数据集引用（<owner>/<name>[@version]）：格式在本地就校验，
+    # 存在性与可见性由服务端在提交准入时裁决。
+    for flag, ref in (("--train-dataset", train_dataset), ("--validation-dataset", validation_dataset)):
+        ds_id = ref.strip().partition("@")[0]
+        if ref.strip() and ("/" not in ds_id or len([s for s in ds_id.split("/") if s]) != 2):
+            raise SpecError(
+                f"{flag} 必须是 <owner>/<name>[@version]（如 alice/gsm8k@v1），收到 {ref!r}；"
+                "用 `sf dataset ls` 查看完整 ID"
+            )
+
+    hyperparams = parse_set(sets or [])
+    if validate:
+        # 与服务端同一份 recipe 声明：不会出现「本地说没问题、服务端说不行」。
+        hyperparams = r.validate_hyperparams(hyperparams, strict=True)
+
+    pool_objs = tuple(parse_pool(p) for p in (pools or []))
+    if not pool_objs:
+        # CLI 总会由 --profile 物化出资源池（materialize_pools），走到这里
+        # 只能是直接调用 build_spec 却漏传 pools。
+        raise SpecError("缺少资源池：请传 pools（CLI 由 --profile 名称[:总卡数] 自动物化）")
+    role_map: dict[str, str] = {}
+    for raw in roles or []:
+        if ":" not in raw:
+            raise SpecError(f"角色映射需要 role:pool 形式，收到 {raw!r}")
+        role, _, pool = raw.partition(":")
+        role_map[role.strip()] = pool.strip()
+
+    if pool_objs and not role_map:
+        # 只有一个池时可以合理推断：recipe 声明的角色全落在它上面。
+        if len(pool_objs) == 1:
+            role_map = {role: pool_objs[0].name for role in r.roles}
+        else:
+            raise SpecError(
+                f"声明了 {len(pool_objs)} 个资源池但缺少角色映射；"
+                f"多池请用 --profile role=名称[:总卡数] 声明。方法 {r.id} 的角色：{', '.join(r.roles)}"
+            )
+    if validate:
+        known = {p.name for p in pool_objs}
+        for role, pool in role_map.items():
+            if role not in r.roles:
+                raise SpecError(f"方法 {r.id} 没有角色 {role!r}；可用：{', '.join(r.roles)}")
+            if pool not in known:
+                raise SpecError(f"角色 {role!r} 指向未定义的资源池 {pool!r}")
+
+    actions = tuple(a.strip() for a in (on_success or []) if a.strip())
+    if validate:
+        for a in actions:
+            if not r.supports(a):
+                raise SpecError(
+                    f"方法 {r.id} 不支持训练后动作 {a!r}；支持：{', '.join(r.lifecycle) or '（无）'}"
+                )
+
+    prov = provenance or {}
+    return JobSpec(
+        metadata=Metadata(
+            name=exp_basename(exp_rel),
+            project=project or "",
+            owner=user or "",          # 服务端会权威覆写
+            display_name=display_name or "",
+        ),
+        spec=JobSpecBody(
+            recipe=RecipeRef(name=r.id, version=r.version, digest=r.digest, plugins=r.plugins),
+            source=SourceSpec(exp=exp_rel),
+            framework=FrameworkRef(
+                kind=r.framework,
+                version=selected_framework_version,
+                runtime_id=selected_runtime.runtime_id,
+                image=selected_image,
+                observability_url=observability_url.strip(),
+            ),
+            model=ModelSpec(
+                base=base_model.strip(),
+                init_from=ArtifactRef.parse(init_from) if init_from else None,
+            ),
+            # dataset 与 path 是两个正交维度：dataset 触发分发（拉到共享缓存、
+            # 注入 <NAME>_DATA_DIR 环境变量），path 是训练框架实际读的位置
+            # （nemo-rl 通常在 config 里用 ${oc.env:<NAME>_DATA_DIR} 引用，不需要 path）。
+            data=DataSpec(
+                train=(
+                    DataRef(dataset=train_dataset.strip(), path=train_data.strip())
+                    if (train_data.strip() or train_dataset.strip()) else None
+                ),
+                validation=(
+                    DataRef(dataset=validation_dataset.strip(), path=validation_data.strip())
+                    if (validation_data.strip() or validation_dataset.strip()) else None
+                ),
+            ),
+            resources=ResourceSpec(pools=pool_objs, roles=role_map),
+            hyperparams=hyperparams,
+            lifecycle=LifecycleSpec(on_success=actions),
+            # 平台托管插件包的精确引用（来自实验的 plugins.lock.json）。
+            # PluginUse.from_dict 顺带做格式校验：id 两段式、digest 必须 sha256:。
+            plugins=tuple(
+                PluginUse.from_dict(p, field_path=f"spec.plugins[{i}]")
+                for i, p in enumerate(plugin_uses or [])
+            ),
+        ),
+        provenance=Provenance(
+            core_version=CORE_VERSION,
+            git_commit=str(prov.get("git_commit") or ""),
+            git_dirty=bool(prov.get("git_dirty")),
+            config_sha=str(prov.get("config_sha") or ""),
+        ),
+    )
+
+
+def infer_recipe(exp_path: Path) -> str:
+    """从实验目录读取显式 recipe：`recipe.lock.json` 的 recipe.name 是唯一事实源。
+
+    锁文件本来就是实验必备（`sf new` 生成、提交前校验 digest），recipe 名与版本
+    都在里面，不再要求单独的 method 标注文件；旧实验遗留的 `method` 文件作为
+    兼容回退仍可读。读不到返回空串，由调用方明确报错；不按其他文件的存在性推断。
+    """
+    lock = exp_path / "recipe.lock.json"
+    if lock.is_file():
+        try:
+            payload = json.loads(lock.read_text(encoding="utf-8"))
+        except (OSError, json.JSONDecodeError):
+            payload = {}
+        name = str((payload.get("recipe") or {}).get("name") or "").strip()
+        if name:
+            return name
+    legacy = exp_path / "method"
+    if legacy.is_file():
+        return legacy.read_text(encoding="utf-8").strip()
+    return ""

starforge_cli-0.1.6.dist-info/METADATA

@@ -0,0 +1,40 @@
+Metadata-Version: 2.5
+Name: starforge-cli
+Version: 0.1.6
+Summary: StarForge（星锻）命令行客户端：pip 安装后用 sf init 创建微调项目，再提交训练。
+Project-URL: Homepage, https://github.com/wccdev/starforge
+Project-URL: Documentation, https://github.com/wccdev/starforge
+Project-URL: Repository, https://github.com/wccdev/starforge
+Keywords: grpo,llm,post-training,rlhf,sft,starforge
+Classifier: Development Status :: 4 - Beta
+Classifier: Environment :: Console
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: datasets>=2.18
+Requires-Dist: pyyaml>=6
+Requires-Dist: starforge-core==0.1.6
+Requires-Dist: typer>=0.12
+Provides-Extra: dev
+Requires-Dist: fastapi>=0.110; extra == 'dev'
+Requires-Dist: httpx>=0.27; extra == 'dev'
+Requires-Dist: pytest>=8; extra == 'dev'
+Requires-Dist: pyyaml>=6; extra == 'dev'
+Requires-Dist: ruff>=0.5; extra == 'dev'
+Description-Content-Type: text/markdown
+
+# starforge-cli
+
+StarForge（星锻）命令行客户端。源码在平台仓 `cli/`，以 PyPI 包 `starforge-cli` 发布，安装后提供 `sf` 命令。
+
+```bash
+pip install starforge-cli
+sf init my-lab --yes
+```
+
+官方示例项目见独立仓库 `starforge-tutorial`。本目录不是微调项目，不要在这里 `sf new`。

starforge_cli-0.1.6.dist-info/RECORD

@@ -0,0 +1,55 @@
+starforge_cli/__init__.py,sha256=bBl2cfgC1O5UwCs3_ba_T_aWkoV4VjViiYfALsE_hDE,129
+starforge_cli/api_client.py,sha256=JL-fou_v8B9ftKpdgehivzLgnqSesftn-IrmwqTy-Dk,25829
+starforge_cli/auth.py,sha256=jdK-1tcGY-xdIZr3FIgCLoUHlCOk6k6hbR3TfOcpy90,13008
+starforge_cli/catalog.py,sha256=aDy3bh_zvPOz70f2goj6Qc-vR8taieYsdUh0kMt-r7s,5750
+starforge_cli/cli.py,sha256=sSxC_zRq5ut-5eYsoAS5Ej27ipnvS1pkcdBPQI-_FHg,3235
+starforge_cli/cli_ui.py,sha256=hHFQe7Ci82n6ZP0Etz4I6I84O_t5aOnb4_xfOXkwjyI,15951
+starforge_cli/client_device.py,sha256=homzz_OOIztOJ8SwgJ1TetnG7Gsz6KiYyTV1RnJ8bKE,3742
+starforge_cli/config_resolve.py,sha256=ln3plTCtp1a0oBgmbCl9DwJ90eAKu-4jD38_6nTfStg,743
+starforge_cli/data_prep.py,sha256=QOo3iWg_UzZyfrH7QaJf1j0KLeY-vlutRHJUlM9hQf8,2157
+starforge_cli/new_experiment.py,sha256=RPQmSSTvY8pi0MSMd-xUyHEOA9ZPTYjgZWtk9EJ_7hk,7305
+starforge_cli/packing.py,sha256=UxqUCe1f3EponHUktrSYZeLTFpWAsUW1Slq0LCHOFrE,7560
+starforge_cli/plugins_lock.py,sha256=tIjDYdu0jBVsCxltSsJTNeiSRtE80JTg7bBqaH_1gmc,2705
+starforge_cli/project.py,sha256=9_DFMG05gQTon3eBnLVYzBkNEAl9B7T05mlF3Ef5SXU,5568
+starforge_cli/recipe_lock.py,sha256=rLMXo-j0mBRzrgw-WQhhZ85l8z0FPCHYc9MZwBNzBr8,17335
+starforge_cli/spec_builder.py,sha256=Qs_ADSvptBEUnbR2biHZIMxwkohGmdChTuGLd55IK5M,15475
+starforge_cli/commands/__init__.py,sha256=xTREWFhmebvT1e--N_DZ3EpbNYhjH0kyuI75WpriOiA,85
+starforge_cli/commands/admin.py,sha256=6VM22LsB4O8wWgeiPFSr-giIKrhZ1273aZdZWl1R03M,6641
+starforge_cli/commands/bench.py,sha256=gqgUZL3wH4GhD6MqPlLMZDlv2oqGQVzjZLuA4uLcaOE,4138
+starforge_cli/commands/common.py,sha256=1JW7VA1c4yf6_vwf2XWhvoqjD4dTjEBE8q2mbWPcnwA,6628
+starforge_cli/commands/dataset.py,sha256=5v8GWoUXbavWkPwvjNOolVQHaWqeIPgVadWuFFiAiu8,6291
+starforge_cli/commands/exp.py,sha256=WDTN91KP5zfTKCs6Hy8wXAFIxB49gUJOUI_k_4e388g,8437
+starforge_cli/commands/init.py,sha256=vw-fCgj-ZfOrxMW-4Kml3dCcX6Pt6oK7gdK-ooVQWWc,2085
+starforge_cli/commands/jobs.py,sha256=fubC0Mr8yRe1mm0DGhDxcKWKa9pKJq7YvmQEo0mTSkE,8886
+starforge_cli/commands/login.py,sha256=CKx19kwyguTY-soWpJN-NmTGaELh2vZhSdU66NuitSY,2172
+starforge_cli/commands/plugin.py,sha256=LOSomV-G3sOVn5LNi3fF9rSC7rSNTS0aOLTdVgMciJY,9732
+starforge_cli/commands/recipe.py,sha256=48998L3lPAsC8NSGZS4gtsXTQCBxNeEhQACZe2kHfII,5444
+starforge_cli/commands/serve.py,sha256=D0W9XWJImYkdCKyC0uN9SjZEWMkjAHz1R7kTeHyqnGs,3245
+starforge_cli/commands/submit.py,sha256=9eyYwvlVg-6WvydEbzfarRq-OlkSP5uDnyjGI4uoO8M,21322
+starforge_cli/commands/sweep.py,sha256=Od6A7Ve6FkIoo3C4v26Apw_UkP4swaxas_ws39Mlawo,6427
+starforge_cli/scaffold/agent-run.py.tmpl,sha256=g0XUzej4p2-5ThcsR5BYpUEBdLjd1_L-zNxWoLeaLQ0,5865
+starforge_cli/scaffold/custom-framework/train.sh,sha256=n2RLgEsk20hcTinaGbvWrXwc9y7xHzxIjlrTChF4j3c,3458
+starforge_cli/scaffold/experiment-template/.gitkeep,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+starforge_cli/scaffold/experiment-template/README.md,sha256=ht-Szx04Fmo70kPkDUmD2-X151fawN9HjFA-rMLUsA0,1184
+starforge_cli/scaffold/experiment-template/config.yaml,sha256=x29cTrWm-VsRYPLqtE5O1BTwm4dEBBwJr5jpeIa4r8M,1663
+starforge_cli/scaffold/project/gitignore,sha256=OF7Hn2W-Ffmw6Uvlblhq-7XoqHd0_JKVq7cYcLZf5rw,108
+starforge_cli/scaffold/project/common/README.md,sha256=6mBNuYYq26A3ck0gqMPF3MEs0xCFdBfL5ZTiqlW0feY,690
+starforge_cli/scaffold/project/common/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+starforge_cli/scaffold/project/configs/README.md,sha256=J8hLHgzpE1q1RN9fRqt_TAFGTKXvAKzhzY5h_apt2dc,5463
+starforge_cli/scaffold/project/configs/base/README.md,sha256=-mFyhx-DEaKOJkyVHYMPIYsAR73Ky4nTi_wJxC2xKyU,1241
+starforge_cli/scaffold/project/configs/base/distillation_math.yaml,sha256=Tj8TI2JjL9CVnEX8KoRlx7-MqbT7H_mTjItYpXWq5U8,10351
+starforge_cli/scaffold/project/configs/base/grpo_lora.yaml,sha256=noVhbwE5HoXNZVdysasKLbw6IPitF8nQz6XmKwnyO6M,1206
+starforge_cli/scaffold/project/configs/base/grpo_math_1B.yaml,sha256=UH68TavoA8TEb1mrUJSoxCXRipmpnVIhSIoU_9atoQk,20156
+starforge_cli/scaffold/project/configs/base/grpo_megatron.yaml,sha256=gdD4OY42GPbiFxVU30HuFjsG75AeS052-bzOGhVqnXA,2398
+starforge_cli/scaffold/project/configs/base/grpo_noncolocated.yaml,sha256=a5Qr6yAQ-9j8_SF5C0R4f3Rtj9aOHJlw6VQ5UEH9ip0,1035
+starforge_cli/scaffold/project/configs/base/grpo_sliding_puzzle.yaml,sha256=GpL4IR4ponTrMA9_N41Q_uKbBYPBkUnvpJB-Urll5Ws,2463
+starforge_cli/scaffold/project/configs/base/ppo_math_1B.yaml,sha256=zsVXO17SzPia1SBdRdrPe5JLrueOaf0kIPlV2r-M4GQ,12845
+starforge_cli/scaffold/project/configs/base/rm.yaml,sha256=mqyi24MmOlMPqvkW_M7W5K3CTDapHE1wrE8Qd2DCVk8,8284
+starforge_cli/scaffold/project/configs/base/sft.yaml,sha256=lr_bnzcNobzWwoOLYI9i-53K4MJ9RlxHt9zXywb6VoI,11162
+starforge_cli/scaffold/project/configs/models/README.md,sha256=bXsDjKpwLcUScXe1xAqce4nv1S8lScTY7Tahu5Jt7gw,778
+starforge_cli/scaffold/project/configs/models/qwen3.5-4b.yaml,sha256=FCDitwTu_xCeawB9_BFM3Y1N0uryq4zowdpR_hY6cxc,478
+starforge_cli/scaffold/project/configs/models/qwen3.5-9b.yaml,sha256=-_o-zn4XlIQVR9dzw_a1g4nFWrvw0P_EY0vPzSEamw0,587
+starforge_cli-0.1.6.dist-info/METADATA,sha256=GnMSEnDFpflGjq1ASKY95IEeQE8lkrmusP8L5aZugwY,1601
+starforge_cli-0.1.6.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
+starforge_cli-0.1.6.dist-info/entry_points.txt,sha256=cH1c_LcMPSnRb3UBTb05zk3TtmFtpvSsVy4ODkh9ctQ,45
+starforge_cli-0.1.6.dist-info/RECORD,,

starforge_cli-0.1.6.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.32.0
+Root-Is-Purelib: true
+Tag: py3-none-any

starforge_cli-0.1.6.dist-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+sf = starforge_cli.cli:app