sft-cli 0.2.0__tar.gz → 0.2.1__tar.gz

{sft_cli-0.2.0 → sft_cli-0.2.1}/.github/workflows/ci.yml

@@ -38,7 +38,7 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-latest, macos-latest]
-        python-version: ["3.9", "3.11", "3.12"]
+        python-version: ["3.9", "3.13", "3.14"]
     steps:
       - uses: actions/checkout@v4
 
@@ -61,7 +61,7 @@ jobs:
     # Long-running tests (subprocess E2E, perf smoke, build-hook) only need
     # to pass on one cell — they exercise behavior that's OS- and version-
     # independent.
-    name: slow tests (py3.12 / ubuntu)
+    name: slow tests (py3.14 / ubuntu)
     runs-on: ubuntu-latest
     needs: test
     steps:
@@ -71,7 +71,7 @@ jobs:
         uses: astral-sh/setup-uv@v4
 
       - name: Set up Python
-        run: uv python install 3.12
+        run: uv python install 3.14
 
       - name: Install dependencies
         run: uv sync --dev
@@ -93,7 +93,7 @@ jobs:
         uses: astral-sh/setup-uv@v4
 
       - name: Set up Python
-        run: uv python install 3.12
+        run: uv python install 3.14
 
       - name: Build package
         run: uv build

{sft_cli-0.2.0 → sft_cli-0.2.1}/.github/workflows/publish.yml

@@ -14,7 +14,7 @@ jobs:
         uses: astral-sh/setup-uv@v4
 
       - name: Set up Python
-        run: uv python install 3.12
+        run: uv python install 3.14
 
       - name: Build package
         run: uv build

sft_cli-0.2.1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Matan Ben-Yosef
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

sft_cli-0.2.1/PKG-INFO

@@ -0,0 +1,155 @@
+Metadata-Version: 2.4
+Name: sft-cli
+Version: 0.2.1
+Summary: An interactive terminal browser for .safetensors files
+Project-URL: Homepage, https://github.com/matanby/sft-cli
+Project-URL: Repository, https://github.com/matanby/sft-cli
+Author: Matan Ben-Yosef
+License: MIT
+License-File: LICENSE
+Keywords: browser,cli,machine-learning,safetensors,tui
+Classifier: Development Status :: 3 - Alpha
+Classifier: Environment :: Console
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Utilities
+Requires-Python: >=3.9
+Requires-Dist: ml-dtypes>=0.3
+Requires-Dist: numpy>=1.24
+Requires-Dist: safetensors>=0.4
+Requires-Dist: textual>=0.40
+Requires-Dist: typer>=0.9
+Provides-Extra: torch
+Requires-Dist: torch>=2.0; extra == 'torch'
+Description-Content-Type: text/markdown
+
+# sft
+
+> The Swiss army knife for `.safetensors` files.
+
+[![PyPI](https://img.shields.io/pypi/v/sft-cli?style=flat-square)](https://pypi.org/project/sft-cli/)
+[![Python](https://img.shields.io/pypi/pyversions/sft-cli?style=flat-square)](https://pypi.org/project/sft-cli/)
+[![CI](https://img.shields.io/github/actions/workflow/status/matanby/sft-cli/ci.yml?branch=main&style=flat-square)](https://github.com/matanby/sft-cli/actions)
+[![License](https://img.shields.io/badge/license-MIT-blue?style=flat-square)](LICENSE)
+
+<img src="https://vhs.charm.sh/vhs-3nFXGLuvC5swABYCewgchD.gif" alt="sft demo">
+
+`sft` is a single-binary CLI for inspecting, editing, and diffing `.safetensors` files — and an interactive terminal browser for poking around large checkpoints. Most commands read the file header only, so multi-gigabyte models open in milliseconds.
+
+It also ships a [skill](#-ai-agents) that teaches AI coding agents (Claude Code, Cursor, Codex CLI) when to reach for it and how to parse the output.
+
+## Install
+
+```bash
+uv tool install sft-cli                 # recommended
+pip install sft-cli                     # or pip
+uv tool install 'sft-cli[torch]'        # + .pt/.pth conversion
+```
+
+## 🚀 Quick start
+
+```bash
+sft model.safetensors                   # open the interactive browser
+sft info model.safetensors              # one-shot summary
+sft info model.safetensors --json       # machine-readable
+```
+
+The bare `sft <file>` form is a shortcut for `sft browse <file>`. Inside the browser: `↑↓` to navigate, `/` to filter, `L` on a LoRA file for LoRA Mode, `D` to diff against another file, `q` to quit.
+
+## What it does
+
+**Inspect** a file without loading it:
+
+```bash
+sft info  model.safetensors             # size, tensor count, dtypes, metadata
+sft ls    model.safetensors             # flat list, sort/filter friendly
+sft tree  model.safetensors --depth=2   # hierarchical view
+sft stat  model.safetensors             # per-tensor mean/std/min/max/sparsity
+sft check model.safetensors             # corruption + NaN/Inf scan
+```
+
+**Compare** two checkpoints:
+
+```bash
+sft diff base.safetensors finetuned.safetensors --delta \
+    --include='**.self_attn.**'         # cosine, L2, max-abs per tensor
+```
+
+**Edit** without writing Python:
+
+```bash
+sft slice  big.safetensors --include='**.weight' -o weights-only.safetensors
+sft strip  big.safetensors --exclude='*lora_*'
+sft cast   model.safetensors --dtype fp16
+sft cat    a.safetensors b.safetensors -o merged.safetensors
+sft rename model.safetensors --sub 'model\.' 'backbone.'
+sft split  model.safetensors --max-size 4GB
+sft convert pytorch_model.bin           # → safetensors
+```
+
+Every write command supports `--dry-run` and never overwrites the input — outputs default to `{stem}.{suffix}.safetensors`.
+
+**Adapter workflows** (PEFT and Kohya):
+
+```bash
+sft lora info    adapter.safetensors            # rank, alpha, target modules
+sft lora svd     adapter.safetensors            # singular-value spectrum
+sft lora compat  base.safetensors adapter.safetensors
+sft lora extract base.safetensors ft.safetensors --rank 16
+sft lora resize  adapter.safetensors --rank auto    # per-pair adaptive rank
+sft lora stack   a.safetensors b.safetensors -a 0.7 -b 0.3
+sft lora merge   base.safetensors adapter.safetensors
+sft lora convert adapter.safetensors --to peft      # Kohya ↔ PEFT
+```
+
+`--rank auto` picks each pair's output rank from its singular-value spectrum (`ceil(stable_rank) + 1`), so over-parameterized pairs compress harder than rich ones. `auto+N` adds a safety margin.
+
+## The browser
+
+Press a key, get a result.
+
+| Key | Action |
+|:---:|---|
+| `↑` `↓` | Navigate |
+| `←` `→` | Collapse / expand tree |
+| `Tab` | Switch between tree and table |
+| `/` | Search / filter |
+| `s` | Cycle sort |
+| `Enter` | Tensor stats popup |
+| `m` | File metadata |
+| `c` | Cast file dtype |
+| `L` | LoRA Mode (per-pair stats, SVD, compress) |
+| `D` | Diff against another file |
+| `:` | Command palette |
+| `q` | Quit |
+
+## 🤖 AI agents
+
+```bash
+sft skill install                       # auto-detects Claude / Cursor / Codex
+sft skill status
+sft skill uninstall
+```
+
+The installer symlinks `sft`'s skill into your agent's well-known skills directory (`~/.claude/skills/sft`, `~/.cursor/skills/sft`, etc.), so it stays in sync when you `uv tool upgrade sft-cli`. Pass `--mode copy` for a frozen snapshot.
+
+Every command supports `--json` for clean parsing:
+
+```bash
+sft info model.safetensors --json | jq '.tensors'
+sft lora info adapter.safetensors --json | jq '.rank'
+sft stat model.safetensors --json --include='**.q_proj.*'
+```
+
+## License
+
+MIT — see [LICENSE](LICENSE).

sft_cli-0.2.1/README.md

@@ -0,0 +1,121 @@
+# sft
+
+> The Swiss army knife for `.safetensors` files.
+
+[![PyPI](https://img.shields.io/pypi/v/sft-cli?style=flat-square)](https://pypi.org/project/sft-cli/)
+[![Python](https://img.shields.io/pypi/pyversions/sft-cli?style=flat-square)](https://pypi.org/project/sft-cli/)
+[![CI](https://img.shields.io/github/actions/workflow/status/matanby/sft-cli/ci.yml?branch=main&style=flat-square)](https://github.com/matanby/sft-cli/actions)
+[![License](https://img.shields.io/badge/license-MIT-blue?style=flat-square)](LICENSE)
+
+<img src="https://vhs.charm.sh/vhs-3nFXGLuvC5swABYCewgchD.gif" alt="sft demo">
+
+`sft` is a single-binary CLI for inspecting, editing, and diffing `.safetensors` files — and an interactive terminal browser for poking around large checkpoints. Most commands read the file header only, so multi-gigabyte models open in milliseconds.
+
+It also ships a [skill](#-ai-agents) that teaches AI coding agents (Claude Code, Cursor, Codex CLI) when to reach for it and how to parse the output.
+
+## Install
+
+```bash
+uv tool install sft-cli                 # recommended
+pip install sft-cli                     # or pip
+uv tool install 'sft-cli[torch]'        # + .pt/.pth conversion
+```
+
+## 🚀 Quick start
+
+```bash
+sft model.safetensors                   # open the interactive browser
+sft info model.safetensors              # one-shot summary
+sft info model.safetensors --json       # machine-readable
+```
+
+The bare `sft <file>` form is a shortcut for `sft browse <file>`. Inside the browser: `↑↓` to navigate, `/` to filter, `L` on a LoRA file for LoRA Mode, `D` to diff against another file, `q` to quit.
+
+## What it does
+
+**Inspect** a file without loading it:
+
+```bash
+sft info  model.safetensors             # size, tensor count, dtypes, metadata
+sft ls    model.safetensors             # flat list, sort/filter friendly
+sft tree  model.safetensors --depth=2   # hierarchical view
+sft stat  model.safetensors             # per-tensor mean/std/min/max/sparsity
+sft check model.safetensors             # corruption + NaN/Inf scan
+```
+
+**Compare** two checkpoints:
+
+```bash
+sft diff base.safetensors finetuned.safetensors --delta \
+    --include='**.self_attn.**'         # cosine, L2, max-abs per tensor
+```
+
+**Edit** without writing Python:
+
+```bash
+sft slice  big.safetensors --include='**.weight' -o weights-only.safetensors
+sft strip  big.safetensors --exclude='*lora_*'
+sft cast   model.safetensors --dtype fp16
+sft cat    a.safetensors b.safetensors -o merged.safetensors
+sft rename model.safetensors --sub 'model\.' 'backbone.'
+sft split  model.safetensors --max-size 4GB
+sft convert pytorch_model.bin           # → safetensors
+```
+
+Every write command supports `--dry-run` and never overwrites the input — outputs default to `{stem}.{suffix}.safetensors`.
+
+**Adapter workflows** (PEFT and Kohya):
+
+```bash
+sft lora info    adapter.safetensors            # rank, alpha, target modules
+sft lora svd     adapter.safetensors            # singular-value spectrum
+sft lora compat  base.safetensors adapter.safetensors
+sft lora extract base.safetensors ft.safetensors --rank 16
+sft lora resize  adapter.safetensors --rank auto    # per-pair adaptive rank
+sft lora stack   a.safetensors b.safetensors -a 0.7 -b 0.3
+sft lora merge   base.safetensors adapter.safetensors
+sft lora convert adapter.safetensors --to peft      # Kohya ↔ PEFT
+```
+
+`--rank auto` picks each pair's output rank from its singular-value spectrum (`ceil(stable_rank) + 1`), so over-parameterized pairs compress harder than rich ones. `auto+N` adds a safety margin.
+
+## The browser
+
+Press a key, get a result.
+
+| Key | Action |
+|:---:|---|
+| `↑` `↓` | Navigate |
+| `←` `→` | Collapse / expand tree |
+| `Tab` | Switch between tree and table |
+| `/` | Search / filter |
+| `s` | Cycle sort |
+| `Enter` | Tensor stats popup |
+| `m` | File metadata |
+| `c` | Cast file dtype |
+| `L` | LoRA Mode (per-pair stats, SVD, compress) |
+| `D` | Diff against another file |
+| `:` | Command palette |
+| `q` | Quit |
+
+## 🤖 AI agents
+
+```bash
+sft skill install                       # auto-detects Claude / Cursor / Codex
+sft skill status
+sft skill uninstall
+```
+
+The installer symlinks `sft`'s skill into your agent's well-known skills directory (`~/.claude/skills/sft`, `~/.cursor/skills/sft`, etc.), so it stays in sync when you `uv tool upgrade sft-cli`. Pass `--mode copy` for a frozen snapshot.
+
+Every command supports `--json` for clean parsing:
+
+```bash
+sft info model.safetensors --json | jq '.tensors'
+sft lora info adapter.safetensors --json | jq '.rank'
+sft stat model.safetensors --json --include='**.q_proj.*'
+```
+
+## License
+
+MIT — see [LICENSE](LICENSE).

{sft_cli-0.2.0 → sft_cli-0.2.1}/pyproject.toml

@@ -20,6 +20,8 @@ classifiers = [
     "Programming Language :: Python :: 3.10",
     "Programming Language :: Python :: 3.11",
     "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Programming Language :: Python :: 3.14",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
     "Topic :: Utilities",
 ]

{sft_cli-0.2.0 → sft_cli-0.2.1}/src/sft/__init__.py

@@ -1,3 +1,3 @@
 """sft — An interactive terminal browser for .safetensors files."""
 
-__version__ = "0.2.0"
+__version__ = "0.2.1"

{sft_cli-0.2.0 → sft_cli-0.2.1}/src/sft/browser.py

@@ -740,6 +740,8 @@ class LoraSortMode(Enum):
     RANK_ASC = "rank ↑"
     EFF_RANK_DESC = "eff. rank ↓"
     EFF_RANK_ASC = "eff. rank ↑"
+    SV95_DESC = "sv95 ↓"
+    SV95_ASC = "sv95 ↑"
     NORM_A_DESC = "‖A‖ ↓"
     NORM_A_ASC = "‖A‖ ↑"
     NORM_B_DESC = "‖B‖ ↓"
@@ -753,6 +755,8 @@ LORA_SORT_ORDER = [
     LoraSortMode.RANK_ASC,
     LoraSortMode.EFF_RANK_DESC,
     LoraSortMode.EFF_RANK_ASC,
+    LoraSortMode.SV95_DESC,
+    LoraSortMode.SV95_ASC,
     LoraSortMode.NORM_A_DESC,
     LoraSortMode.NORM_A_ASC,
     LoraSortMode.NORM_B_DESC,
@@ -1295,6 +1299,7 @@ class LoraModeScreen(Screen):
         self.lora_info = lora_info  # only populated for PEFT
         self._sort_idx = 0
         self._stats: dict[str, dict[str, float]] = {}
+        self._tensor_cache: dict | None = None
 
     def compose(self) -> ComposeResult:
         yield Footer()
@@ -1331,12 +1336,26 @@ class LoraModeScreen(Screen):
         LoraSortMode.RANK_ASC: ("rank", " ↑"),
         LoraSortMode.EFF_RANK_DESC: ("eff_rank", " ↓"),
         LoraSortMode.EFF_RANK_ASC: ("eff_rank", " ↑"),
+        LoraSortMode.SV95_DESC: ("sv95", " ↓"),
+        LoraSortMode.SV95_ASC: ("sv95", " ↑"),
         LoraSortMode.NORM_A_DESC: ("norm_a", " ↓"),
         LoraSortMode.NORM_A_ASC: ("norm_a", " ↑"),
         LoraSortMode.NORM_B_DESC: ("norm_b", " ↓"),
         LoraSortMode.NORM_B_ASC: ("norm_b", " ↑"),
     }
 
+    _LORA_COLUMN_SORT: dict[str, tuple[LoraSortMode, LoraSortMode]] = {
+        "module": (LoraSortMode.MODULE_ASC, LoraSortMode.MODULE_DESC),
+        "rank": (LoraSortMode.RANK_ASC, LoraSortMode.RANK_DESC),
+        "eff_rank": (LoraSortMode.EFF_RANK_ASC, LoraSortMode.EFF_RANK_DESC),
+        "sv95": (LoraSortMode.SV95_ASC, LoraSortMode.SV95_DESC),
+        "norm_a": (LoraSortMode.NORM_A_ASC, LoraSortMode.NORM_A_DESC),
+        "norm_b": (LoraSortMode.NORM_B_ASC, LoraSortMode.NORM_B_DESC),
+    }
+
+    _SORT_MISSING_ASC = float("inf")
+    _SORT_MISSING_DESC = float("-inf")
+
     def on_mount(self) -> None:
         if self.lora_format == "peft" and self.lora_info:
             table = self.query_one("#lora-table", DataTable)
@@ -1363,6 +1382,8 @@ class LoraModeScreen(Screen):
             self._compute_stats()
             self._update_sort_indicator()
 
+    _STATS_REFRESH_EVERY = 64
+
     def _title_line(self) -> str:
         fmt_label = self.lora_format.upper()
         return (
@@ -1397,10 +1418,24 @@ class LoraModeScreen(Screen):
                 "…",
                 key=pair.module_key,
             )
+        self._sync_module_column_width(table)
+
+    def _sync_module_column_width(self, table: DataTable) -> None:
+        if self.lora_info is None:
+            return
+        max_len = max(
+            (len(self._short_module_name(p)) for p in self.lora_info.pairs),
+            default=16,
+        )
+        col = table.columns.get("module")
+        if col is not None:
+            col.content_width = max(col.content_width, max_len)
+        table.refresh(layout=True)
 
     def _short_module_name(self, pair: LoRAPair) -> str:
-        parts = pair.module_key.split(".")
-        return ".".join(parts[-4:]) if len(parts) > 4 else pair.module_key
+        from sft.ops.lora.detect import format_lora_module_display
+
+        return format_lora_module_display(pair.module_key)
 
     def _sorted_pairs(self) -> list[LoRAPair]:
         if self.lora_info is None:
@@ -1408,8 +1443,9 @@ class LoraModeScreen(Screen):
         mode = LORA_SORT_ORDER[self._sort_idx]
         pairs = list(self.lora_info.pairs)
 
-        def stat(name: str, key: str, default: float) -> float:
-            return self._stats.get(name, {}).get(key, default)
+        def stat(name: str, key: str, *, ascending: bool) -> float:
+            missing = self._SORT_MISSING_ASC if ascending else self._SORT_MISSING_DESC
+            return self._stats.get(name, {}).get(key, missing)
 
         if mode == LoraSortMode.MODULE_ASC:
             pairs.sort(key=lambda p: natural_sort_key(p.module_key))
@@ -1422,42 +1458,56 @@ class LoraModeScreen(Screen):
         elif mode == LoraSortMode.EFF_RANK_DESC:
             pairs.sort(
                 key=lambda p: (
-                    -stat(p.module_key, "eff_rank", 0.0),
+                    -stat(p.module_key, "eff_rank", ascending=False),
                     natural_sort_key(p.module_key),
                 )
             )
         elif mode == LoraSortMode.EFF_RANK_ASC:
             pairs.sort(
                 key=lambda p: (
-                    stat(p.module_key, "eff_rank", 0.0),
+                    stat(p.module_key, "eff_rank", ascending=True),
+                    natural_sort_key(p.module_key),
+                )
+            )
+        elif mode == LoraSortMode.SV95_DESC:
+            pairs.sort(
+                key=lambda p: (
+                    -stat(p.module_key, "sv95", ascending=False),
+                    natural_sort_key(p.module_key),
+                )
+            )
+        elif mode == LoraSortMode.SV95_ASC:
+            pairs.sort(
+                key=lambda p: (
+                    stat(p.module_key, "sv95", ascending=True),
                     natural_sort_key(p.module_key),
                 )
             )
         elif mode == LoraSortMode.NORM_A_DESC:
             pairs.sort(
                 key=lambda p: (
-                    -stat(p.module_key, "norm_a", 0.0),
+                    -stat(p.module_key, "norm_a", ascending=False),
                     natural_sort_key(p.module_key),
                 )
             )
         elif mode == LoraSortMode.NORM_A_ASC:
             pairs.sort(
                 key=lambda p: (
-                    stat(p.module_key, "norm_a", 0.0),
+                    stat(p.module_key, "norm_a", ascending=True),
                     natural_sort_key(p.module_key),
                 )
             )
         elif mode == LoraSortMode.NORM_B_DESC:
             pairs.sort(
                 key=lambda p: (
-                    -stat(p.module_key, "norm_b", 0.0),
+                    -stat(p.module_key, "norm_b", ascending=False),
                     natural_sort_key(p.module_key),
                 )
             )
         elif mode == LoraSortMode.NORM_B_ASC:
             pairs.sort(
                 key=lambda p: (
-                    stat(p.module_key, "norm_b", 0.0),
+                    stat(p.module_key, "norm_b", ascending=True),
                     natural_sort_key(p.module_key),
                 )
             )
@@ -1509,6 +1559,7 @@ class LoraModeScreen(Screen):
                 new_cursor = i
         if table.row_count > 0:
             table.move_cursor(row=new_cursor)
+        self._sync_module_column_width(table)
         self._update_sort_indicator()
 
     @staticmethod
@@ -1541,17 +1592,20 @@ class LoraModeScreen(Screen):
             )
             return
 
-        for pair in self.lora_info.pairs:
+        self._tensor_cache = tensors
+        total = len(self.lora_info.pairs)
+
+        for i, pair in enumerate(self.lora_info.pairs):
             try:
                 a = tensors[pair.lora_a_name]
                 b = tensors[pair.lora_b_name]
 
                 (s,) = _qr_svd(a, b, compute_uv=False)
                 s_sq = s**2
-                total = float(s_sq.sum())
-                if total > 0:
-                    eff_rank = total / float(s_sq.max())
-                    cumvar = np.cumsum(s_sq) / total
+                total_energy = float(s_sq.sum())
+                if total_energy > 0:
+                    eff_rank = total_energy / float(s_sq.max())
+                    cumvar = np.cumsum(s_sq) / total_energy
                     sv95 = int(np.searchsorted(cumvar, 0.95)) + 1
                 else:
                     eff_rank = 0.0
@@ -1569,19 +1623,43 @@ class LoraModeScreen(Screen):
                 stats = {}
 
             self._stats[pair.module_key] = stats
-            self.app.call_from_thread(self._refresh_table)
+            if (i + 1) % self._STATS_REFRESH_EVERY == 0 or (i + 1) == total:
+                self.app.call_from_thread(self._refresh_table)
 
     # --- Actions ---
 
     def action_exit_mode(self) -> None:
         self.app.pop_screen()
 
+    def _set_sort_mode(self, mode: LoraSortMode) -> None:
+        if self.lora_format != "peft":
+            return
+        self._sort_idx = LORA_SORT_ORDER.index(mode)
+        self._refresh_table()
+
     def action_cycle_sort(self) -> None:
         if self.lora_format != "peft":
             return
         self._sort_idx = (self._sort_idx + 1) % len(LORA_SORT_ORDER)
         self._refresh_table()
 
+    def on_data_table_header_selected(self, event: DataTable.HeaderSelected) -> None:
+        """Click a column header to toggle asc/desc for that metric."""
+        if self.lora_format != "peft" or event.control.id != "lora-table":
+            return
+        col_key = event.column_key.value
+        modes = self._LORA_COLUMN_SORT.get(col_key)
+        if modes is None:
+            return
+        asc, desc = modes
+        current = LORA_SORT_ORDER[self._sort_idx]
+        if current == asc:
+            self._set_sort_mode(desc)
+        elif current == desc:
+            self._set_sort_mode(asc)
+        else:
+            self._set_sort_mode(desc if col_key != "module" else asc)
+
     def on_data_table_row_selected(self, _event: DataTable.RowSelected) -> None:
         self.action_show_spectrum()
 
@@ -1785,8 +1863,9 @@ class LoraInfoScreen(ModalScreen):
 
     @staticmethod
     def _short_pair_name(pair: LoRAPair) -> str:
-        parts = pair.module_key.split(".")
-        return ".".join(parts[-4:]) if len(parts) > 4 else pair.module_key
+        from sft.ops.lora.detect import format_lora_module_display
+
+        return format_lora_module_display(pair.module_key)
 
 
 class LoraResizePromptScreen(ModalScreen):
@@ -1961,8 +2040,11 @@ class SvdSpectrumScreen(ModalScreen):
         self.sv = [float(v) for v in singular_values]
 
     def compose(self) -> ComposeResult:
+        from sft.ops.lora.detect import format_lora_module_display
+
+        module_label = format_lora_module_display(self.pair.module_key)
         with Container(id="svd-container"):
-            yield Label(f"SVD Spectrum — {self.pair.target_module}", id="svd-title")
+            yield Label(f"SVD Spectrum — {module_label}", id="svd-title")
             yield Static(
                 f"[dim]Module:[/dim] {self.pair.module_key}\n"
                 f"[dim]Rank:[/dim] {self.pair.rank}   "

{sft_cli-0.2.0 → sft_cli-0.2.1}/src/sft/commands/lora.py

@@ -115,6 +115,7 @@ def lora_info_cmd(
     """
     file = validate_safetensors(file)
 
+    from sft.ops.lora.detect import format_lora_module_display
     from sft.ops.lora.info import lora_info
 
     try:
@@ -153,7 +154,7 @@ def lora_info_cmd(
 
     typer.echo("Pairs:")
     for pair in info.pairs:
-        module_short = pair.module_key.split(".")[-2] + "." + pair.target_module
+        module_short = format_lora_module_display(pair.module_key)
         a_shape = "x".join(str(d) for d in pair.lora_a_shape)
         b_shape = "x".join(str(d) for d in pair.lora_b_shape)
         typer.echo(
@@ -516,6 +517,7 @@ def lora_svd_cmd(
             "threshold": analysis.threshold,
             "modules": [
                 {
+                    "module_key": m.module_key,
                     "module": m.module,
                     "rank": m.rank,
                     "sv_90": m.sv_90,
@@ -530,11 +532,16 @@ def lora_svd_cmd(
         typer.echo(json.dumps(data, indent=2))
         return
 
-    header = f"{'Module':<16}{'Rank':>6}{'SV 90%':>10}{'SV 95%':>10}{'SV 99%':>10}{'Suggested rank':>16}"
+    col_width = max(16, *(len(m.module) for m in analysis.modules))
+    header = (
+        f"{'Module':<{col_width}}{'Rank':>6}{'SV 90%':>10}{'SV 95%':>10}"
+        f"{'SV 99%':>10}{'Suggested rank':>16}"
+    )
     typer.echo(header)
     for m in analysis.modules:
         typer.echo(
-            f"{m.module:<16}{m.rank:>6}{m.sv_90:>10}{m.sv_95:>10}{m.sv_99:>10}{m.suggested_rank:>16}"
+            f"{m.module:<{col_width}}{m.rank:>6}{m.sv_90:>10}{m.sv_95:>10}"
+            f"{m.sv_99:>10}{m.suggested_rank:>16}"
         )
 
     typer.echo()