1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
use clap::{Args, Subcommand};
use super::util::llm_model_completion_parser;
#[derive(Debug, Args)]
pub(crate) struct ModelsArgs {
#[command(subcommand)]
pub command: ModelsCommand,
}
#[derive(Debug, Subcommand)]
pub(crate) enum ModelsCommand {
/// Print resolved metadata for a model alias or model id as JSON.
Info(ModelInfoArgs),
/// Inspect LoRA adapter metadata and compatibility with a Harn model route.
Lora(ModelsLoraArgs),
/// List models grouped by provider.
List(ModelsListArgs),
/// Pull an Ollama model or print setup steps for a known local runtime.
Install(ModelsInstallArgs),
/// Recommend a starter model for the current machine and credentials.
Recommend(ModelRecommendArgs),
/// Round-trip a small prompt through a model and report timing, tokens, and cost.
Test(ModelsTestArgs),
}
#[derive(Debug, Args)]
pub(crate) struct ModelsLoraArgs {
#[command(subcommand)]
pub command: ModelsLoraCommand,
}
#[derive(Debug, Subcommand)]
pub(crate) enum ModelsLoraCommand {
/// Export a tool-calling corpus into a trainer-ready LoRA dataset.
Export(ModelsLoraExportArgs),
/// Inspect a PEFT LoRA adapter directory or repo id.
Inspect(ModelsLoraInspectArgs),
/// Plan a portable LoRA/QLoRA tool-calling fine-tune for a Harn model route.
Plan(ModelsLoraPlanArgs),
/// Check a tool-calling corpus before spending GPU time on LoRA training.
Preflight(ModelsLoraPreflightArgs),
}
#[derive(Debug, Args)]
pub(crate) struct ModelsLoraExportArgs {
/// Base model alias or provider-native id the dataset targets.
#[arg(long = "base", value_parser = llm_model_completion_parser(), hide_possible_values = true)]
pub base_model: String,
/// Provider/runtime to plan against instead of inferring from the base model.
#[arg(long)]
pub provider: Option<String>,
/// Tool-call format to export for (`auto`, `native`, `text`, or `json`).
#[arg(long = "tool-format", default_value = "auto")]
pub tool_format: String,
/// Corpus JSONL file, or a directory containing a conventional corpus JSONL.
#[arg(long, value_name = "PATH")]
pub corpus: String,
/// Write exported JSONL rows to this path. Required unless --check is set.
#[arg(long, value_name = "PATH")]
pub out: Option<std::path::PathBuf>,
/// Write a provenance manifest with input/output hashes and export stats.
#[arg(long, value_name = "PATH")]
pub manifest: Option<std::path::PathBuf>,
/// Validate conversion and print a report without writing dataset rows.
#[arg(long)]
pub check: bool,
/// Served LoRA adapter/model name to include in row metadata.
#[arg(long = "adapter-name")]
pub adapter_name: Option<String>,
/// Chat template identifier to include in row metadata.
#[arg(long = "chat-template")]
pub chat_template: Option<String>,
/// Extra target provenance copied into row metadata, as KEY=VALUE.
#[arg(long = "target-metadata", value_name = "KEY=VALUE")]
pub target_metadata: Vec<String>,
/// Emit structured JSON report.
#[arg(long)]
pub json: bool,
}
#[derive(Debug, Args)]
pub(crate) struct ModelsLoraInspectArgs {
/// Base model alias or provider-native id the adapter will attach to.
#[arg(long = "base", value_parser = llm_model_completion_parser(), hide_possible_values = true)]
pub base_model: String,
/// Adapter directory or Hugging Face repo id.
pub adapter: String,
/// Request model name to expose for the adapter. Defaults to the adapter directory/repo basename.
#[arg(long)]
pub name: Option<String>,
/// Provider/runtime to check against instead of inferring from the base model.
#[arg(long)]
pub provider: Option<String>,
/// LoRA export manifest to compare against the adapter and requested route.
#[arg(long, value_name = "PATH")]
pub manifest: Option<std::path::PathBuf>,
/// Fail when the adapter config omits or mismatches the manifest contract id.
#[arg(long = "require-contract-id")]
pub require_contract_id: bool,
/// Emit structured JSON.
#[arg(long)]
pub json: bool,
}
#[derive(Debug, Args)]
pub(crate) struct ModelsLoraPlanArgs {
/// Base model alias or provider-native id to fine-tune.
#[arg(long = "base", value_parser = llm_model_completion_parser(), hide_possible_values = true)]
pub base_model: String,
/// Provider/runtime to plan against instead of inferring from the base model.
#[arg(long)]
pub provider: Option<String>,
/// Tool-call format to train for (`auto`, `native`, `text`, or `json`).
#[arg(long = "tool-format", default_value = "auto")]
pub tool_format: String,
/// Corpus path or dataset id to include in the generated eval/training recipe.
#[arg(long, value_name = "PATH_OR_DATASET")]
pub corpus: Option<String>,
/// Optional teacher model route for synthetic corpus refresh or distillation.
#[arg(long, value_parser = llm_model_completion_parser(), hide_possible_values = true)]
pub teacher: Option<String>,
/// Corpus strategy (`auto`, `audit-only`, `refresh`, or `distill`).
#[arg(long = "corpus-strategy", default_value = "auto")]
pub corpus_strategy: String,
/// Adapter training method (`qlora` or `lora`).
#[arg(long, default_value = "qlora")]
pub method: String,
/// LoRA rank to plan for training and serving.
#[arg(long, default_value_t = 16)]
pub rank: u32,
/// LoRA alpha. Defaults to 2 * --rank.
#[arg(long)]
pub alpha: Option<u32>,
/// LoRA dropout probability.
#[arg(long, default_value_t = 0.05)]
pub dropout: f64,
/// Emit structured JSON.
#[arg(long)]
pub json: bool,
}
#[derive(Debug, Args)]
pub(crate) struct ModelsLoraPreflightArgs {
/// Base model alias or provider-native id the adapter will target.
#[arg(long = "base", value_parser = llm_model_completion_parser(), hide_possible_values = true)]
pub base_model: String,
/// Provider/runtime to check against instead of inferring from the base model.
#[arg(long)]
pub provider: Option<String>,
/// Corpus JSONL file, or a directory containing a conventional corpus JSONL.
#[arg(long, value_name = "PATH")]
pub corpus: String,
/// Training config file to read max_seq_length/min_fit_ratio from.
#[arg(long, value_name = "PATH")]
pub config: Option<std::path::PathBuf>,
/// Override the config's max_seq_length.
#[arg(long = "max-seq-length")]
pub max_seq_length: Option<u64>,
/// Override the config's min_fit_ratio.
#[arg(long = "min-fit-ratio")]
pub min_fit_ratio: Option<f64>,
/// Hard approximate token budget outlier ceiling.
#[arg(long = "hard-token-limit", default_value_t = 32_768)]
pub hard_token_limit: u64,
/// Minimum trainable record count.
#[arg(long = "min-records", default_value_t = 1)]
pub min_records: u64,
/// Expected source tool-call body format (`json`, `text`, or `auto`).
#[arg(long = "source-tool-format", default_value = "json")]
pub source_tool_format: String,
/// Minimum share of tool calls matching --source-tool-format.
#[arg(long = "min-tool-call-share", default_value_t = 0.95)]
pub min_tool_call_share: f64,
/// Require the last assistant message in each trainable record to contain this marker.
#[arg(long = "done-marker")]
pub done_marker: Option<String>,
/// Exit non-zero when readiness checks fail.
#[arg(long)]
pub check: bool,
/// Emit structured JSON.
#[arg(long)]
pub json: bool,
}
#[derive(Debug, Args)]
pub(crate) struct ModelInfoArgs {
/// Verify provider-local readiness for the resolved model when supported.
#[arg(long)]
pub verify: bool,
/// Warm/preload the resolved model when supported. Implies --verify.
#[arg(long)]
pub warm: bool,
/// Ollama keep_alive value to use with --warm (for example 30m, forever, or -1).
#[arg(long = "keep-alive", value_name = "VALUE")]
pub keep_alive: Option<String>,
/// Model alias or provider-native model id.
#[arg(
value_parser = llm_model_completion_parser(),
hide_possible_values = true
)]
pub model: String,
}
#[derive(Debug, Args)]
pub(crate) struct ModelsListArgs {
/// Restrict to a single provider.
#[arg(long)]
pub provider: Option<String>,
/// Emit JSON instead of a human table.
#[arg(long)]
pub json: bool,
/// Only show locally-installed (Ollama) models.
#[arg(long = "installed-only")]
pub installed_only: bool,
}
#[derive(Debug, Args)]
pub(crate) struct ModelsInstallArgs {
/// Model alias or provider-native id to install or set up.
pub model: String,
/// Skip the size-confirmation prompt.
#[arg(long)]
pub yes: bool,
/// Optional Ollama keep-alive hint (e.g. `5m`, `1h`).
#[arg(long = "keep-alive", value_name = "VALUE")]
pub keep_alive: Option<String>,
}
#[derive(Debug, Args)]
pub(crate) struct ModelRecommendArgs {
/// Emit the recommendation and hardware snapshot as JSON.
#[arg(long)]
pub json: bool,
}
#[derive(Debug, Args)]
pub(crate) struct ModelsTestArgs {
/// Model alias or provider-native model id.
pub model: String,
/// Prompt text to send to the model.
#[arg(long, default_value = "Reply with the word pong.")]
pub prompt: String,
/// Provider id to use instead of inferring one from the model selector.
#[arg(long)]
pub provider: Option<String>,
/// Emit a structured JSON result.
#[arg(long, default_value_t = false)]
pub json: bool,
}