Skip to content

armgit / Soup

Public Imported from GitHub
0

Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.

refactor(export): remove deprecated awq and gptq export formats (#1710) (#1951)

Nariman Abdikarim5a6df6b63bbc
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 14130820..ee54b967 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -266,7 +266,7 @@ an issue number or a version. Existing test files keep their names.
| test_multi_adapter.py | Multi-adapter serving: validation, parsing, FastAPI endpoints, CLI |
| test_data_sample.py | Data sampling: random/diverse/hard strategies, CLI, edge cases |
| test_adapters.py | Adapter management: list/info/compare, discovery, metadata |
-| test_awq_gptq_export.py | AWQ/GPTQ export: format support, CLI, quantize mocks, calibration, security |
+| test_export_removed_formats.py | AWQ/GPTQ removal verification: rejection at CLI, pyproject extras, codebase clean |
| test_packing.py | Sample packing: config, YAML, trainer integration, sweep |
| test_data_split.py | Data split: ratio/absolute/stratified splits, seed, edge cases |
| test_curriculum.py | Curriculum learning: config, length sort, buckets, sweep |
diff --git a/README.ar.md b/README.ar.md
index 0a358f38..19096307 100644
--- a/README.ar.md
+++ b/README.ar.md
@@ -1,4 +1,4 @@
-<!-- synced-from: README.md sha256:4675902b8ea66e77e0822a60fdc9d3205a11eb02fb70cdf3b5670bd9ef1af391 -->
+<!-- synced-from: README.md sha256:b10e7bad5b4287de70b59bd0776c162f21da7e7cc0e2eda004dfdfc2bcc7fd24 -->
<p align="center">🌍 <a href="README.md">English</a> | <a href="README.tr.md">Türkçe</a> | <strong>العربية</strong> | <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a></p>
<p align="center">
@@ -135,9 +135,6 @@ soup train
ثنائي الجانب متينًا إزاء التباين، ويوقّع `soup adapters sign --backend sigstore`
و`soup attest emit --sign sigstore` باستخدام Sigstore OIDC بلا مفاتيح
(<span dir="ltr">`pip install "soup-cli[sigstore]"`</span>).
-- **قيد الإيقاف التدريجي (Deprecated):** سيُزال `soup export --format awq` و<span dir="ltr">`--format gptq`</span>
- في v0.77؛ فكلا المشروعين الأصليين مؤرشف ولا يُثبَّت أي من الملحقين (extras) إلى جانب
- <span dir="ltr">`[train]`</span>.
> Python **3.11–3.13** فقط. وعلى الإصدار 3.14 وما بعده كان pip يختار حزم PyTorch (wheels) غير مختبرة تنهار في
> الامتداد الأصلي قبل أن يعمل Soup أصلًا.
@@ -216,7 +213,7 @@ soup merge --adapter ./output # ادمج LoRA في
soup export --model ./output --format gguf --quant q4_k_m # GGUF لـ Ollama / llama.cpp
```
-وتوجد أهداف تصدير أخرى (ONNX وTensorRT وAWQ وGPTQ وBitNet) وخيارات النشر في
+وتوجد أهداف تصدير أخرى (ONNX وTensorRT وBitNet) وخيارات النشر في
[`docs/serving-and-export.md`](docs/serving-and-export.md).
## واجهة الويب
diff --git a/README.ja.md b/README.ja.md
index afb0cf1c..f1e1b0de 100644
--- a/README.ja.md
+++ b/README.ja.md
@@ -1,4 +1,4 @@
-<!-- synced-from: README.md sha256:4675902b8ea66e77e0822a60fdc9d3205a11eb02fb70cdf3b5670bd9ef1af391 -->
+<!-- synced-from: README.md sha256:b10e7bad5b4287de70b59bd0776c162f21da7e7cc0e2eda004dfdfc2bcc7fd24 -->
<p align="center">🌍 <a href="README.md">English</a> | <a href="README.tr.md">Türkçe</a> | <a href="README.ar.md">العربية</a> | <strong>日本語</strong> | <a href="README.zh.md">中文</a></p>
<p align="center">
@@ -135,8 +135,6 @@ LLM の学習は、今でも骨の折れる作業です。経験豊富なチー
使わなかった最も速い実行より 2.9x 速くデコードできました)。`soup ab` は、分散に頑健な両側検定を実行します。
`soup adapters sign --backend sigstore` と `soup attest emit --sign sigstore` は、Sigstore のキーレス OIDC で
署名します(`pip install "soup-cli[sigstore]"`)。
-- **非推奨:** `soup export --format awq` と `--format gptq` は v0.77 で削除されます。どちらの
- 上流プロジェクトもアーカイブ済みで、どちらの extra も `[train]` と一緒にはインストールできません。
> Python は **3.11–3.13** のみ対応です。3.14 以降では、pip がテストされていない PyTorch の wheel を解決し、
> Soup が動き出す前にネイティブ拡張の中でクラッシュしていました。
@@ -217,7 +215,7 @@ soup merge --adapter ./output # LoRA をベース
soup export --model ./output --format gguf --quant q4_k_m # Ollama / llama.cpp 向けの GGUF
```
-その他のエクスポート先(ONNX、TensorRT、AWQ、GPTQ、BitNet)とデプロイのオプションは、
+その他のエクスポート先(ONNX、TensorRT、BitNet)とデプロイのオプションは、
[`docs/serving-and-export.md`](docs/serving-and-export.md) にあります。
## Web UI
diff --git a/README.md b/README.md
index 1ccd8110..38c36a6c 100644
--- a/README.md
+++ b/README.md
@@ -136,8 +136,6 @@ that will refuse them. A config that loaded before can stop loading: run
2.9x faster than the fastest without). `soup ab` runs a two-sided, variance-robust test, and
`soup adapters sign --backend sigstore` and `soup attest emit --sign sigstore` sign with
Sigstore keyless OIDC (`pip install "soup-cli[sigstore]"`).
-- **Deprecated:** `soup export --format awq` and `--format gptq` will be removed in v0.77; both
- upstream projects are archived and neither extra installs next to `[train]`.
> From v0.77, Python **3.11–3.13** only: 3.10 reached end of life on 2026-10-01 (pip there stays
> on v0.76); on 3.14+, pip falls back to v0.72.4 (no upper bound) and untested PyTorch wheels.
@@ -218,7 +216,7 @@ soup merge --adapter ./output # merge LoRA into th
soup export --model ./output --format gguf --quant q4_k_m # GGUF for Ollama / llama.cpp
```
-More export targets (ONNX, TensorRT, AWQ, GPTQ, BitNet) and deployment options live in
+More export targets (ONNX, TensorRT, BitNet) and deployment options live in
[`docs/serving-and-export.md`](docs/serving-and-export.md).
## Web UI
diff --git a/README.tr.md b/README.tr.md
index 9539f755..7e6b59b2 100644
--- a/README.tr.md
+++ b/README.tr.md
@@ -1,4 +1,4 @@
-<!-- synced-from: README.md sha256:4675902b8ea66e77e0822a60fdc9d3205a11eb02fb70cdf3b5670bd9ef1af391 -->
+<!-- synced-from: README.md sha256:b10e7bad5b4287de70b59bd0776c162f21da7e7cc0e2eda004dfdfc2bcc7fd24 -->
<p align="center">🌍 <a href="README.md">English</a> | <strong>Türkçe</strong> | <a href="README.ar.md">العربية</a> | <a href="README.ja.md">日本語</a> | <a href="README.zh.md">中文</a></p>
<p align="center">
@@ -138,8 +138,6 @@ veriyor ve onları reddedecek sürümü belirtiyor. Daha önce yüklenen bir yap
kod çözmede 2.9x daha hızlıydı). `soup ab`, iki yönlü, varyansa karşı dayanıklı bir test çalıştırır;
`soup adapters sign --backend sigstore` ve `soup attest emit --sign sigstore` ise Sigstore anahtarsız
OIDC ile imzalar (`pip install "soup-cli[sigstore]"`).
-- **Kullanımdan kaldırıldı:** `soup export --format awq` ve `--format gptq` v0.77'de kaldırılacak; her
- iki üst proje de arşivlenmiş durumda ve ekstralardan hiçbiri `[train]` ile yan yana kurulamıyor.
> v0.77'den itibaren yalnızca Python **3.11–3.13**. Python 3.10, 2026-10-01'de kullanım ömrünün
> sonuna (end of life) ulaştı; 3.10 üzerinde pip v0.76'yı kurmaya devam eder. 3.14+ sürümlerinde
@@ -222,7 +220,7 @@ soup merge --adapter ./output # LoRA'yı tabana bi
soup export --model ./output --format gguf --quant q4_k_m # Ollama / llama.cpp için GGUF
```
-Diğer dışa aktarma hedefleri (ONNX, TensorRT, AWQ, GPTQ, BitNet) ve dağıtım seçenekleri
+Diğer dışa aktarma hedefleri (ONNX, TensorRT, BitNet) ve dağıtım seçenekleri
[`docs/serving-and-export.md`](docs/serving-and-export.md) içinde yer alır.
## Web Arayüzü
diff --git a/README.zh.md b/README.zh.md
index 7d3bba9a..8bcf1180 100644
--- a/README.zh.md
+++ b/README.zh.md
@@ -1,4 +1,4 @@
-<!-- synced-from: README.md sha256:4675902b8ea66e77e0822a60fdc9d3205a11eb02fb70cdf3b5670bd9ef1af391 -->
+<!-- synced-from: README.md sha256:b10e7bad5b4287de70b59bd0776c162f21da7e7cc0e2eda004dfdfc2bcc7fd24 -->
<p align="center">🌍 <a href="README.md">English</a> | <a href="README.tr.md">Türkçe</a> | <a href="README.ar.md">العربية</a> | <a href="README.ja.md">日本語</a> | <strong>中文</strong></p>
<p align="center">
@@ -131,8 +131,6 @@ NF4 达到 **952 MHz 下 119.6 tok/s,峰值 3.32 GB**(修复前)——与
`soup ab` 运行双侧、对方差稳健的检验,`soup adapters sign --backend sigstore` 和
`soup attest emit --sign sigstore` 使用 Sigstore 免密钥 OIDC 签名
(`pip install "soup-cli[sigstore]"`)。
-- **已弃用:** `soup export --format awq` 和 `--format gptq` 将在 v0.77 移除;这两个上游项目都已归档,
- 两个 extra 也都无法与 `[train]` 一起安装。
> 仅支持 Python **3.11–3.13**。在 3.14+ 上,pip 过去会解析到未测试过的 PyTorch
> wheel,在 Soup 运行之前就崩在原生扩展里。
@@ -212,7 +210,7 @@ soup merge --adapter ./output # 把 LoRA 合进基
soup export --model ./output --format gguf --quant q4_k_m # 导出 GGUF 给 Ollama / llama.cpp
```
-更多导出目标(ONNX、TensorRT、AWQ、GPTQ、BitNet)和部署选项在
+更多导出目标(ONNX、TensorRT、BitNet)和部署选项在
[`docs/serving-and-export.md`](docs/serving-and-export.md)。
## Web UI
diff --git a/changelog.d/0.76.0/1951.removed.md b/changelog.d/0.76.0/1951.removed.md
new file mode 100644
index 00000000..981c8290
--- /dev/null
+++ b/changelog.d/0.76.0/1951.removed.md
@@ -0,0 +1 @@
+- **Removed deprecated `awq` and `gptq` export and quantize formats and extras (#338, #1690, #1710, in #1951 by @Nari-Ab).** Both upstream projects (`autoawq` and `auto-gptq`) are archived and incompatible with `[train]`. `soup export --format awq` / `--format gptq` and `soup quantize --to awq` / `--to gptq` are now refused at the CLI with non-zero exit codes naming the remaining supported formats (`gguf`, `onnx`, `tensorrt`, `bitnet`, `tq1_0`, `torchao`, `gguf-ud`). `soup export` also lost `--bits`, `--group-size`, and `--calibration-samples`, so scripts passing them now exit 2 with `No such option`; `--calibration-data` is now read only by `--format gguf-ud`. The `[awq]` and `[gptq]` extras and related export helpers are removed in v0.77; users who still need AWQ or GPTQ export should remain on 0.76.x.
diff --git a/docs/backends-and-ops.md b/docs/backends-and-ops.md
index 133aefad..151d640c 100644
--- a/docs/backends-and-ops.md
+++ b/docs/backends-and-ops.md
@@ -945,7 +945,7 @@ the tokenizer still comes from the adapter folder.
are recorded as such; a cache-path hint on any backend is still unverified. Standalone
wrapper/sweep invocations do not automatically establish a run scope. Metadata is
written for the final artifact, not each intermediate checkpoint. Other exporters
-(including ONNX/TensorRT and native AWQ/GPTQ/torchao loaders), remote adapter sidecars
+(including ONNX/TensorRT and native torchao loaders), remote adapter sidecars
and independent teacher/reward/scorer repositories do not gain a binding guarantee.
Evaluation engines that bypass the shared live-eval loader (including lm-eval's
adapter benchmark path and vLLM) also remain unbound in this stage.
diff --git a/docs/cli-reference.md b/docs/cli-reference.md
index 093e220f..62978cf2 100644
--- a/docs/cli-reference.md
+++ b/docs/cli-reference.md
@@ -79,7 +79,7 @@ Usage: `soup [OPTIONS] <command> [ARGS]...`
| [`soup env`](#soup-env) | Hermetic env lockfile + ABI drift detection (v0.64.0). |
| [`soup eval`](#soup-eval) | Evaluate models: benchmarks, custom evals, LLM judge, leaderboard. |
| [`soup expect`](#soup-expect) | Run an expectations suite against a JSONL dataset. |
-| [`soup export`](#soup-export) | Export a model to GGUF, ONNX, TensorRT-LLM, AWQ, GPTQ, or TorchAO format. |
+| [`soup export`](#soup-export) | Export a model to GGUF, ONNX, TensorRT-LLM, BitNet, or TorchAO format. |
| [`soup fetch`](#soup-fetch) | Fetch a ready-to-edit example config from the bundled catalog. |
| [`soup history`](#soup-history) | Show the lineage tree for all entries with a given name. |
| [`soup infer`](#soup-infer) | Run batch inference on a JSONL file of prompts. |
@@ -2328,14 +2328,14 @@ Usage: `soup expect [OPTIONS] DATA SUITE`
## `soup export`
-Export a model to GGUF, ONNX, TensorRT-LLM, AWQ, GPTQ, or TorchAO format.
+Export a model to GGUF, ONNX, TensorRT-LLM, BitNet, or TorchAO format.
Usage: `soup export [OPTIONS]`
| Name | Type | Default | Required | Help |
|---|---|---|---|---|
| `--model`, `-m` | text | | yes | Path to model directory (full model or LoRA adapter) |
-| `--format`, `-f` | text | `gguf` | | Export format: gguf, onnx, tensorrt, awq, gptq, bitnet, tq1_0, torchao, gguf-ud (torchao needs --quant-config; awq and gptq are deprecated and will be removed in Soup v0.77) |
+| `--format`, `-f` | text | `gguf` | | Export format: gguf, onnx, tensorrt, bitnet, tq1_0, torchao, gguf-ud (torchao needs --quant-config) |
| `--quant`, `-q` | text | `q4_k_m` | | Quantization type: q4_0, q4_k_m, q5_k_m, q8_0, f16, f32 |
| `--output`, `-o` | text | | | Output file path. Default: &lt;model-name>.&lt;quant>.gguf |
| `--base`, `-b` | text | | | Base model ID (for LoRA adapters). Auto-detected if not set. |
@@ -2343,10 +2343,7 @@ Usage: `soup export [OPTIONS]`
| `--onnx-task` | text | `text-generation` | | ONNX export task: text-generation (causal LM) or feature-extraction (embedding) |
| `--deploy` | text | | | Auto-deploy after export. Currently supported: ollama |
| `--deploy-name` | text | | | Model name for deployment (used with --deploy) |
-| `--bits` | integer | `4` | | Quantization bits for AWQ/GPTQ: 4 or 8 |
-| `--group-size` | integer | `128` | | Group size for AWQ/GPTQ quantization |
-| `--calibration-data` | text | | | Path to calibration JSONL. Required for AWQ and GPTQ. |
-| `--calibration-samples` | integer | `128` | | Number of calibration samples for AWQ/GPTQ |
+| `--calibration-data` | text | | | Path to calibration JSONL for imatrix GGUF export. |
| `--registry-id` | text | | | Attach exported artifact to this registry entry (default: auto-match by source --model output dir) |
| `--trust-remote-code` | flag | off | | Allow loading models that ship custom Python via auto_map. Default deny (v0.36.0). Only enable if you trust the source. |
| `--quant-config` | text | | | Path to YAML for torchao PTQ export (v0.53.1 #142). Required when --format=torchao. |
@@ -3059,8 +3056,8 @@ Usage: `soup quantize [OPTIONS] MODEL_PATH`
| Name | Type | Default | Required | Help |
|---|---|---|---|---|
| `MODEL_PATH` | text | | yes | Source checkpoint (safetensors directory or single .safetensors). |
-| `--to` | text | `gguf` | | Target format: gguf \| gptq \| awq \| onnx \| tensorrt (awq and gptq are deprecated and will be removed in Soup v0.77). |
-| `--bits` | integer | `4` | | Quantization bits (1-16; respected by gguf/gptq/awq). |
+| `--to` | text | `gguf` | | Target format: gguf \| onnx \| tensorrt. |
+| `--bits` | integer | `4` | | Quantization bits (1-16; respected by gguf). |
| `--output`, `-o` | text | | | Destination directory (default: &lt;model_path>-&lt;to>). |
## `soup quickstart`
diff --git a/docs/commands.md b/docs/commands.md
index 4fba8f86..86296a11 100644
--- a/docs/commands.md
+++ b/docs/commands.md
@@ -46,8 +46,6 @@ soup export --model ./output --format gguf Export to GGUF (Ollama)
soup export --model ./output --deploy ollama Export GGUF + auto-deploy to Ollama
soup export --model ./output --format onnx Export to ONNX
soup export --model ./output --format tensorrt Export to TensorRT-LLM
-soup export --model ./output --format awq --calibration-data cal.jsonl Export to AWQ (4-bit; deprecated, removed in Soup v0.77)
-soup export --model ./output --format gptq --calibration-data cal.jsonl Export to GPTQ (4-bit; deprecated, removed in Soup v0.77)
soup export --model ./output --format gguf --verbose Show export phases and stream converter/quantizer logs while they run (#1760)
soup export --model ./output --format gguf --no-install-deps Never invoke pip; refuse missing conversion packages before merge/toolchain work (#1760)
soup deploy ollama --model m.gguf --name x Deploy GGUF to Ollama
diff --git a/docs/models.md b/docs/models.md
index 5faa926e..5f473d1d 100644
--- a/docs/models.md
+++ b/docs/models.md
@@ -159,11 +159,8 @@ spelling is unchanged. See [the README's install section](../README.md#1-install
| `liger` | `pip install "soup-cli[liger]"` | Liger Kernel fused ops |
| `ring-attn` | `pip install "soup-cli[ring-attn]"` | Ring FlashAttention (sequence parallelism) |
| `onnx` / `tensorrt` | `pip install "soup-cli[onnx]"` | ONNX / TensorRT-LLM export |
-| `awq` / `gptq` | `pip install "soup-cli[awq]"` | AWQ / GPTQ quantized export; not installable alongside `[train]` (see below) |
| `trackers` | `pip install "soup-cli[trackers]"` | MLflow / SwanLab / Trackio logging |
| `remote` | `pip install "soup-cli[remote]"` | Remote datasets (s3 / gs / az / oci) |
| `dev` | `pip install "soup-cli[dev]"` | Tests + lint + types (pytest, ruff, mypy, pre-commit) |
-**`[awq]` and `[gptq]` cannot be installed alongside `[train]`.** AWQ export was measured working only with `transformers` 4.52.4 or older, while `[train]` requires `transformers>=5.16.1`; `auto-gptq` publishes no Python 3.12 wheel; and both upstream projects (AutoAWQ and AutoGPTQ) are archived. Run AWQ or GPTQ export from a separate environment. Both export formats are deprecated and will be removed in Soup v0.77 ([#338](https://github.com/MakazhanAlpamys/Soup/issues/338)); until then `soup export --format awq` and `--format gptq` start with a notice that names the formats to move to (`gguf`, `onnx`, `tensorrt`, `bitnet`, `tq1_0`).
-
The complete, authoritative extras list is in [`pyproject.toml`](../pyproject.toml).
diff --git a/docs/performance-and-quantization.md b/docs/performance-and-quantization.md
index 0f94a0a9..cab7888e 100644
--- a/docs/performance-and-quantization.md
+++ b/docs/performance-and-quantization.md
@@ -1617,12 +1617,6 @@ group_size: 32
EOF
soup export --model ./merged --format torchao --quant-config ./q.yaml --output ./out
-# AWQ/GPTQ export (deprecated, removed in Soup v0.77) — an explicit local calibration set is required
-soup export --model ./merged --format awq \
- --calibration-data ./calib.jsonl --output ./out-awq
-soup export --model ./merged --format gptq \
- --calibration-data ./calib.jsonl --output ./out-gptq
-
# Unsloth Dynamic 2.0 / IQ / Apple-ARM GGUF via llama.cpp imatrix
soup export --model ./merged --format gguf-ud \
--gguf-flavour UD-Q4_K_XL \
@@ -1638,10 +1632,6 @@ soup deploy autopilot --target rtx-4090-24gb \
Autopilot also detects pre-quantized bases automatically - `TheBloke/Llama-2-7B-Chat-GPTQ` is recommended `gptq` instead of stacking 4-bit on top. Detection runs against the base-model name regex AND any local `config.json`'s `quantization_config.quant_method`. Out-of-cwd model paths are silently skipped (soft-probe semantics). A detected `mxfp4` base is budgeted as the bf16 model it loads (2 bytes per parameter), and Autopilot refuses a `--gpu-budget` below that.
-AWQ and GPTQ export are deprecated and will be removed in Soup v0.77: both upstream projects (AutoAWQ and AutoGPTQ) are archived and neither extra can be installed next to `[train]`. Until then `soup export --format awq` and `--format gptq` work as before and start with a `Deprecated:` notice that names the formats to move to (`gguf`, `onnx`, `tensorrt`, `bitnet`, `tq1_0`).
-
-Direct AWQ and GPTQ exports require `--calibration-data`. Soup refuses a missing or unusable JSONL before importing the quantizer or loading the model. This keeps calibration inputs explicit and prevents AutoAWQ from silently downloading its large default dataset. Use `--calibration-samples` to cap the number of usable JSONL rows (default: 128).
-
The advanced GGUF pipeline uses POSIX `O_NOFOLLOW` to defeat the TOCTOU race between the dispatch-time symlink check and the actual open of the calibration data — a crafted environment cannot race-swap the calibration file between validate and read.
`soup deploy autopilot --measure` caches results at `~/.soup/deploy_autopilot_cache.json` keyed on `(base, profile, eval-tasks)`. Repeat invocations short-circuit; pass `SOUP_DEPLOY_AUTOPILOT_CACHE=<path>` to redirect (constrained to home / cwd / tempdir). The recommended candidate uses soft-fallback: first `OK` by insertion order, else the candidate with the smallest delta (least drop relative to its own baseline).
diff --git a/docs/serving-and-export.md b/docs/serving-and-export.md
index 2668cbe1..20709e10 100644
--- a/docs/serving-and-export.md
+++ b/docs/serving-and-export.md
@@ -2,7 +2,7 @@
[← Back to the Soup README](../README.md)
-> The OpenAI-compatible inference server, batch inference, benchmarking, merge/export (GGUF/ONNX/TensorRT/AWQ/GPTQ), the Anthropic Messages endpoint, speculative decoding, deploy autopilot, the Web UI, and Agent Forge.
+> The OpenAI-compatible inference server, batch inference, benchmarking, merge/export (GGUF/ONNX/TensorRT/BitNet/TorchAO), the Anthropic Messages endpoint, speculative decoding, deploy autopilot, the Web UI, and Agent Forge.
**Contents:**
@@ -247,8 +247,8 @@ Auto-detected chat templates: `chatml`, `llama`, `mistral`, `vicuna`, `zephyr` (
`soup export` integrates with Soup's local artifact registry and deployment workflows:
-- **Deploying on export (`--deploy`)**: `--deploy ollama` is supported for GGUF-family formats (`gguf`, `gguf-ud`, `bitnet`, `tq1_0`). For non-GGUF formats (`onnx`, `tensorrt`, `awq`, `gptq`, `torchao`), `--deploy` fails fast with an informative error because Ollama cannot load non-GGUF weight files. Use `--deploy-name` to assign a custom model name in Ollama.
-- **Registering exports (`--registry-id`)**: Pass `--registry-id <id>` to record and attach exported artifacts to a local registry entry across all export formats (`gguf`, `gguf-ud`, `bitnet`, `tq1_0`, `onnx`, `tensorrt`, `awq`, `gptq`, `torchao`). When an export produces a directory (such as multi-shard safetensors, ONNX data, or TensorRT engine files in `<output>/engine/*.engine`), all produced weight and engine files are discovered and attached to the registry entry.
+- **Deploying on export (`--deploy`)**: `--deploy ollama` is supported for GGUF-family formats (`gguf`, `gguf-ud`, `bitnet`, `tq1_0`). For non-GGUF formats (`onnx`, `tensorrt`, `torchao`), `--deploy` fails fast with an informative error because Ollama cannot load non-GGUF weight files. Use `--deploy-name` to assign a custom model name in Ollama.
+- **Registering exports (`--registry-id`)**: Pass `--registry-id <id>` to record and attach exported artifacts to a local registry entry across all export formats (`gguf`, `gguf-ud`, `bitnet`, `tq1_0`, `onnx`, `tensorrt`, `torchao`). When an export produces a directory (such as multi-shard safetensors, ONNX data, or TensorRT engine files in `<output>/engine/*.engine`), all produced weight and engine files are discovered and attached to the registry entry.
## Batch Inference
@@ -814,7 +814,7 @@ soup quantize ./out --to gguf --bits 4
soup quantize ./out --to onnx -o ./out-onnx
```
-Prints the equivalent `soup export …` invocation (escaped via `shlex.quote`) for copy-paste. Intentionally does NOT in-process call `soup export` — Typer commands aren't safe to re-enter. `--to awq` and `--to gptq` still print their command, followed by a notice that the two formats are deprecated and will be removed in Soup v0.77.
+Prints the equivalent `soup export …` invocation (escaped via `shlex.quote`) for copy-paste. Intentionally does NOT in-process call `soup export` — Typer commands aren't safe to re-enter.
## Llama.cpp Proxy
diff --git a/pyproject.toml b/pyproject.toml
index 7fe802af..1ee47a1b 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -182,8 +182,6 @@ ring-attn = ["ring-flash-attn>=0.1.0"]
onnx = ["optimum[onnxruntime]>=1.16.0"]
tensorrt = ["tensorrt_llm>=0.9.0"]
audio = ["librosa>=0.10.0", "soundfile>=0.12.0", "torchaudio>=2.6.0,<3.0.0"]
-awq = ["autoawq>=0.2.0"]
-gptq = ["auto-gptq>=0.7.0"]
sglang = ["sglang>=0.5.1", "fastapi>=0.104.0", "uvicorn>=0.24.0", "httpx>=0.24.0"]
mlx = [
"mlx>=0.20.0",
diff --git a/src/soup_cli/commands/export.py b/src/soup_cli/commands/export.py
index 2da74faa..75992cba 100644
--- a/src/soup_cli/commands/export.py
+++ b/src/soup_cli/commands/export.py
@@ -20,7 +20,7 @@ from soup_cli.utils.terminal import for_terminal, strip_control
console = Console()
SUPPORTED_FORMATS = (
- "gguf", "onnx", "tensorrt", "awq", "gptq",
+ "gguf", "onnx", "tensorrt",
# v0.52.0 Part D — BitNet 1.58-bit + TQ1_0 GGUF.
# Live llama.cpp conversion (v0.71.20 #134); needs a built llama.cpp toolchain.
"bitnet", "tq1_0",
@@ -30,43 +30,12 @@ SUPPORTED_FORMATS = (
# v0.53.1 #139 — UD/IQ/Apple-ARM GGUFs via llama.cpp imatrix.
"gguf-ud",
)
-# #338 — AWQ and GPTQ export are deprecated: both upstream projects are archived
-# and neither extra installs next to the train extra. The two formats still run
-# for one more release; the notice names the formats to move to. Maps the format
-# to the upstream project the notice names.
-DEPRECATED_FORMATS = {"awq": "AutoAWQ", "gptq": "AutoGPTQ"}
-REPLACEMENT_FORMATS = ("gguf", "onnx", "tensorrt", "bitnet", "tq1_0")
GGUF_QUANT_TYPES = ("q4_0", "q4_k_m", "q5_k_m", "q8_0", "f16", "f32")
LLAMA_CPP_DIR_NAME = "llama.cpp"
# Pin to a known release tag for supply-chain safety
LLAMA_CPP_TAG = "b5270"
-def deprecated_format_notice(fmt: str) -> Optional[str]:
- """The deprecation line for an export format, or ``None`` when it stays.
-
- Plain text with no console markup: ``soup export`` and ``soup quantize``
- both print it behind a ``Deprecated:`` label.
- """
- upstream = DEPRECATED_FORMATS.get(fmt)
- if upstream is None:
- return None
- replacements = ", ".join(REPLACEMENT_FORMATS[:-1]) + f" or {REPLACEMENT_FORMATS[-1]}"
- return (
- f"--format {fmt} is deprecated and will be removed in "
- f"Soup v{DEPRECATED_VALUE_REJECTION_VERSION}: "
- f"{upstream} is archived upstream and the {fmt} extra cannot be installed "
- f"next to the train extra. Use --format {replacements} instead."
- )
-
-
-def _warn_deprecated_format(fmt: str) -> None:
- """Print the deprecation notice for ``fmt``, if it has one."""
- notice = deprecated_format_notice(fmt)
- if notice is not None:
- console.print(f"[yellow]Deprecated:[/] {notice}")
-
-
def export(
model: str = typer.Option(
...,
@@ -79,9 +48,8 @@ def export(
"--format",
"-f",
help=(
- "Export format: gguf, onnx, tensorrt, awq, gptq, bitnet, tq1_0, torchao, "
- "gguf-ud (torchao needs --quant-config; awq and gptq are deprecated and "
- f"will be removed in Soup v{DEPRECATED_VALUE_REJECTION_VERSION})"
+ "Export format: gguf, onnx, tensorrt, bitnet, tq1_0, torchao, "
+ "gguf-ud (torchao needs --quant-config)"
),
),
quant: str = typer.Option(
@@ -122,25 +90,10 @@ def export(
"--deploy-name",
help="Model name for deployment (used with --deploy)",
),
- bits: int = typer.Option(
- 4,
- "--bits",
- help="Quantization bits for AWQ/GPTQ: 4 or 8",
- ),
- group_size: int = typer.Option(
- 128,
- "--group-size",
- help="Group size for AWQ/GPTQ quantization",
- ),
calibration_data: Optional[str] = typer.Option(
None,
"--calibration-data",
- help="Path to calibration JSONL. Required for AWQ and GPTQ.",
- ),
- calibration_samples: int = typer.Option(
- 128,
- "--calibration-samples",
- help="Number of calibration samples for AWQ/GPTQ",
+ help="Path to calibration JSONL for imatrix GGUF export.",
),
registry_id: Optional[str] = typer.Option(
None,
@@ -188,7 +141,7 @@ def export(
help="GGUF only: never invoke pip; refuse missing conversion dependencies before work.",
),
):
- """Export a model to GGUF, ONNX, TensorRT-LLM, AWQ, GPTQ, or TorchAO format."""
+ """Export a model to GGUF, ONNX, TensorRT-LLM, BitNet, or TorchAO format."""
# v0.53.10 #152 — pre-fetch the base model from a non-HF hub. ``model``
# is typically a local merged dir / adapter dir; only ``base`` is rewritten.
if hub and hub != "hf":
@@ -206,17 +159,22 @@ def export(
model_path = Path(model)
# --- Validate ---
- if not model_path.exists():
- console.print(f"[red]Model path not found: {model_path}[/]")
- raise typer.Exit(1)
-
if fmt not in SUPPORTED_FORMATS:
+ removed_notice = ""
+ if fmt in ("awq", "gptq"):
+ removed_notice = (
+ f"\n[yellow]--format {fmt} was removed in v{DEPRECATED_VALUE_REJECTION_VERSION}.[/]"
+ )
console.print(
- f"[red]Unsupported format: {fmt}[/]\n"
+ f"[red]Unsupported format: {fmt}[/]{removed_notice}\n"
f"Supported: {', '.join(SUPPORTED_FORMATS)}"
)
raise typer.Exit(1)
+ if not model_path.exists():
+ console.print(f"[red]Model path not found: {model_path}[/]")
+ raise typer.Exit(1)
+
if (verbose or no_install_deps) and fmt != "gguf":
flags = [name for name, enabled in (
("--verbose", verbose), ("--no-install-deps", no_install_deps),
@@ -261,32 +219,6 @@ def export(
)
return
- # --- AWQ export path ---
- if fmt == "awq":
- output_path = _export_awq(
- model_path, output, base, bits, group_size,
- calibration_data, calibration_samples, trust_remote_code,
- )
- if output_path is not None:
- _maybe_attach_export(
- artifact_path=str(output_path), kind="awq",
- explicit_id=registry_id, source_model=str(Path(model)),
- )
- return
-
- # --- GPTQ export path ---
- if fmt == "gptq":
- output_path = _export_gptq(
- model_path, output, base, bits, group_size,
- calibration_data, calibration_samples, trust_remote_code,
- )
- if output_path is not None:
- _maybe_attach_export(
- artifact_path=str(output_path), kind="gptq",
- explicit_id=registry_id, source_model=str(Path(model)),
- )
- return
-
# --- TorchAO PTQ export path (v0.53.1 #142) ---
if fmt == "torchao":
output_path = _export_torchao_cli(
@@ -1087,385 +1019,6 @@ def _export_tensorrt(
return output_path
-def _validate_output_path(output: Optional[str]) -> Optional[Path]:
- """Validate output path stays under cwd (path traversal protection)."""
- if output is None:
- return None
- # realpath + commonpath containment (is_under_cwd) — Path.resolve() +
- # relative_to() breaks on Windows 8.3 short names.
- from soup_cli.utils.paths import is_under_cwd
-
- out_path = Path(output).resolve()
- if not is_under_cwd(output):
- console.print("[red]Output path must be under the current working directory.[/]")
- raise typer.Exit(1)
- return out_path
-
-
-def _validate_calibration_path(calibration_data: Optional[str]) -> Optional[Path]:
- """Validate calibration data path stays under cwd."""
- if calibration_data is None:
- return None
- from soup_cli.utils.paths import is_under_cwd
-
- cal_path = Path(calibration_data).resolve()
- if not is_under_cwd(calibration_data):
- console.print("[red]Calibration data path must be under the current working directory.[/]")
- raise typer.Exit(1)
- if not cal_path.exists():
- console.print(f"[red]Calibration data not found: {cal_path}[/]")
- raise typer.Exit(1)
- return cal_path
-
-
-class _CalibrationDataReadError(ValueError):
- """A calibration JSONL could not be decoded or read safely."""
-
-
-def _load_calibration_texts(cal_path: Optional[Path], max_samples: int = 128) -> list:
- """Load calibration texts from JSONL file."""
- if cal_path is None:
- return []
- texts = []
- try:
- with open(cal_path, encoding="utf-8") as fh:
- for line in fh:
- line = line.strip()
- if not line:
- continue
- try:
- row = json.loads(line)
- except json.JSONDecodeError:
- continue
- if not isinstance(row, dict):
- continue
- # Support "text" field or concatenate all non-empty values.
- if "text" in row:
- value = row["text"]
- text = "" if value is None else str(value)
- else:
- text = " ".join(str(v) for v in row.values() if v)
- if text.strip():
- texts.append(text)
- if len(texts) >= max_samples:
- break
- except UnicodeError as exc:
- raise _CalibrationDataReadError(
- f"Calibration data is not valid UTF-8: {cal_path}"
- ) from exc
- except OSError as exc:
- raise _CalibrationDataReadError(
- f"Could not read calibration data {cal_path}: {exc}"
- ) from exc
- return texts
-
-
-def _standardize_gptq_shard_name(output_path: Path, bits: int, group_size: int) -> None:
- """auto_gptq's save_quantized names its shard gptq_model-<bits>bit-<group>g.safetensors,
- a name AutoModelForCausalLM.from_pretrained does not look for; rename it to the
- standard model.safetensors so the exported directory reloads."""
- shard = output_path / f"gptq_model-{bits}bit-{group_size}g.safetensors"
- standard = output_path / "model.safetensors"
- if shard.exists() and not standard.exists():
- shard.rename(standard)
-
-
-def _export_awq(
- model_path: Path,
- output: Optional[str],
- base: Optional[str],
- bits: int = 4,
- group_size: int = 128,
- calibration_data: Optional[str] = None,
- calibration_samples: int = 128,
- trust_remote_code: bool = False,
-) -> Optional[Path]:
- """Export model to AWQ format via autoawq."""
- # First, so it is there whatever stops the run (most stop at the extra).
- _warn_deprecated_format("awq")
- # Validate bits
- valid_bits = {4, 8}
- if bits not in valid_bits:
- console.print(
- f"[red]Invalid --bits {bits}. Must be one of: {sorted(valid_bits)}[/]"
- )
- raise typer.Exit(1)
-
- # Validate output path (security: path traversal protection)
- validated_output = _validate_output_path(output)
-
- # Validate calibration path (security: path traversal protection)
- cal_path = _validate_calibration_path(calibration_data)
- if cal_path is None:
- console.print(
- "[red]AWQ export requires --calibration-data.[/]\n"
- "AutoAWQ otherwise downloads its large default calibration dataset: "
- "pass a calibration JSONL, e.g. [bold]--calibration-data path/to/data.jsonl[/]."
- )
- raise typer.Exit(1)
- try:
- calib_data = _load_calibration_texts(cal_path, max_samples=calibration_samples)
- except _CalibrationDataReadError as exc:
- console.print(f"[red]{exc}[/]")
- raise typer.Exit(1) from exc
- if not calib_data:
- console.print(f"[red]No usable calibration samples found in {cal_path}.[/]")
- raise typer.Exit(1)
-
- try:
- from awq import AutoAWQForCausalLM
- except ImportError as exc:
- if exc.name != "awq":
- console.print(f"[red]autoawq import failed:[/] {exc}")
- raise typer.Exit(1)
- console.print(
- "[red]autoawq not installed.[/]\n"
- "Install with: [bold]pip install \"soup-cli\\[awq]\"[/]\n"
- "Or directly: [bold]pip install autoawq[/]"
- )
- raise typer.Exit(1)
-
- from transformers import AutoTokenizer
-
- # Check if LoRA adapter — merge first
- adapter_config_path = model_path / "adapter_config.json"
- is_adapter = adapter_config_path.exists()
- merge_dir = None
- source_path = model_path
-
- if is_adapter:
- console.print("[yellow]LoRA adapter detected - merging with base model first...[/]")
- base_model = base or _detect_base_model(adapter_config_path)
- if not base_model:
- console.print(
- "[red]Cannot detect base model from adapter_config.json.[/]\n"
- "Please specify with [bold]--base[/] flag."
- )
- raise typer.Exit(1)
- merge_dir = model_path.parent / f".soup_merge_tmp_{model_path.name}"
- _merge_adapter(str(model_path), base_model, str(merge_dir), trust_remote_code)
- source_path = merge_dir
-
- default_out = model_path.parent / f"{model_path.name}_awq"
- output_path = validated_output if validated_output else default_out
-
- console.print(
- Panel(
- f"Model: [bold]{source_path}[/]\n"
- f"Format: [bold]AWQ[/]\n"
- f"Bits: [bold]{bits}[/]\n"
- f"Group size: [bold]{group_size}[/]\n"
- f"Output: [bold]{output_path}[/]",
- title="Export Plan",
- )
- )
-
- try:
- console.print(
- Panel(
- "[yellow]Warning:[/] Loading model with trust_remote_code=True.\n"
- "This may execute custom code from the model directory.",
- title="Security Notice",
- )
- )
- console.print("[dim]Loading model for AWQ quantization...[/]")
- model = AutoAWQForCausalLM.from_pretrained(str(source_path))
- from soup_cli.utils.trust_remote import (
- model_requires_trust_remote_code,
- resolve_trust_remote_code,
- )
-
- requires_tok = model_requires_trust_remote_code(str(source_path)) or False
- trc_tok = resolve_trust_remote_code(
- str(source_path),
- requested=trust_remote_code,
- console=console,
- requires_remote_code=requires_tok,
- )
- tokenizer = AutoTokenizer.from_pretrained(str(source_path), trust_remote_code=trc_tok)
-
- quant_config = {"zero_point": True, "q_group_size": group_size, "w_bit": bits}
-
- console.print(f"[dim]Quantizing to AWQ {bits}-bit (group_size={group_size})...[/]")
- model.quantize(tokenizer, quant_config=quant_config, calib_data=calib_data)
-
- console.print("[dim]Saving quantized model...[/]")
- model.save_quantized(str(output_path))
- tokenizer.save_pretrained(str(output_path))
-
- except Exception as exc:
- console.print(f"[red]AWQ export failed:[/] {exc}")
- raise typer.Exit(1)
- finally:
- if merge_dir and merge_dir.exists():
- console.print("[dim]Cleaning up temporary merge files...[/]")
- shutil.rmtree(merge_dir, ignore_errors=True)
-
- console.print(
- Panel(
- f"Output: [bold]{output_path}[/]\n"
- f"Format: [bold]AWQ {bits}-bit[/]\n\n"
- f"Use with vLLM:\n"
- f" [bold]from vllm import LLM[/]\n"
- f" [bold]llm = LLM(model='{output_path}', quantization='awq')[/]",
- title="[bold green]AWQ Export Complete![/]",
- )
- )
- return output_path
-
-
-def _export_gptq(
- model_path: Path,
- output: Optional[str],
- base: Optional[str],
- bits: int = 4,
- group_size: int = 128,
- calibration_data: Optional[str] = None,
- calibration_samples: int = 128,
- trust_remote_code: bool = False,
-) -> Optional[Path]:
- """Export model to GPTQ format via auto-gptq."""
- # First, so it is there whatever stops the run (most stop at the extra).
- _warn_deprecated_format("gptq")
- # Validate bits
- valid_bits = {4, 8}
- if bits not in valid_bits:
- console.print(
- f"[red]Invalid --bits {bits}. Must be one of: {sorted(valid_bits)}[/]"
- )
- raise typer.Exit(1)
-
- # Validate output path (security: path traversal protection)
- validated_output = _validate_output_path(output)
-
- # Validate calibration path (security: path traversal protection)
- cal_path = _validate_calibration_path(calibration_data)
- if cal_path is None:
- console.print(
- "[red]GPTQ export requires --calibration-data.[/]\n"
- "auto-gptq has no built-in calibration dataset: "
- "pass a calibration JSONL, e.g. [bold]--calibration-data path/to/data.jsonl[/]."
- )
- raise typer.Exit(1)
- try:
- calib_texts = _load_calibration_texts(cal_path, max_samples=calibration_samples)
- except _CalibrationDataReadError as exc:
- console.print(f"[red]{exc}[/]")
- raise typer.Exit(1) from exc
- if not calib_texts:
- console.print(f"[red]No usable calibration samples found in {cal_path}.[/]")
- raise typer.Exit(1)
-
- try:
- from auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig
- except ImportError as exc:
- if exc.name != "auto_gptq":
- console.print(f"[red]auto-gptq import failed:[/] {exc}")
- raise typer.Exit(1)
- console.print(
- "[red]auto-gptq not installed.[/]\n"
- "Install with: [bold]pip install \"soup-cli\\[gptq]\"[/]\n"
- "Or directly: [bold]pip install auto-gptq[/]"
- )
- raise typer.Exit(1)
-
- from transformers import AutoTokenizer
-
- # Check if LoRA adapter — merge first
- adapter_config_path = model_path / "adapter_config.json"
- is_adapter = adapter_config_path.exists()
- merge_dir = None
- source_path = model_path
-
- if is_adapter:
- console.print("[yellow]LoRA adapter detected - merging with base model first...[/]")
- base_model = base or _detect_base_model(adapter_config_path)
- if not base_model:
- console.print(
- "[red]Cannot detect base model from adapter_config.json.[/]\n"
- "Please specify with [bold]--base[/] flag."
- )
- raise typer.Exit(1)
- merge_dir = model_path.parent / f".soup_merge_tmp_{model_path.name}"
- _merge_adapter(str(model_path), base_model, str(merge_dir), trust_remote_code)
- source_path = merge_dir
-
- default_out = model_path.parent / f"{model_path.name}_gptq"
- output_path = validated_output if validated_output else default_out
-
- console.print(
- Panel(
- f"Model: [bold]{source_path}[/]\n"
- f"Format: [bold]GPTQ[/]\n"
- f"Bits: [bold]{bits}[/]\n"
- f"Group size: [bold]{group_size}[/]\n"
- f"Output: [bold]{output_path}[/]",
- title="Export Plan",
- )
- )
-
- try:
- console.print(
- Panel(
- "[yellow]Warning:[/] Loading model with trust_remote_code=True.\n"
- "This may execute custom code from the model directory.",
- title="Security Notice",
- )
- )
- console.print("[dim]Loading model for GPTQ quantization...[/]")
- quantize_config = BaseQuantizeConfig(
- bits=bits,
- group_size=group_size,
- desc_act=False,
- )
- model = AutoGPTQForCausalLM.from_pretrained(
- str(source_path), quantize_config=quantize_config
- )
- from soup_cli.utils.trust_remote import (
- model_requires_trust_remote_code,
- resolve_trust_remote_code,
- )
-
- requires_tok = model_requires_trust_remote_code(str(source_path)) or False
- trc_tok = resolve_trust_remote_code(
- str(source_path),
- requested=trust_remote_code,
- console=console,
- requires_remote_code=requires_tok,
- )
- tokenizer = AutoTokenizer.from_pretrained(str(source_path), trust_remote_code=trc_tok)
-
- calib_data = [tokenizer(t, return_tensors="pt") for t in calib_texts]
-
- console.print(f"[dim]Quantizing to GPTQ {bits}-bit (group_size={group_size})...[/]")
- model.quantize(calib_data)
-
- console.print("[dim]Saving quantized model...[/]")
- model.save_quantized(str(output_path))
- tokenizer.save_pretrained(str(output_path))
- _standardize_gptq_shard_name(output_path, bits, group_size)
-
- except Exception as exc:
- console.print(f"[red]GPTQ export failed:[/] {exc}")
- raise typer.Exit(1)
- finally:
- if merge_dir and merge_dir.exists():
- console.print("[dim]Cleaning up temporary merge files...[/]")
- shutil.rmtree(merge_dir, ignore_errors=True)
-
- console.print(
- Panel(
- f"Output: [bold]{output_path}[/]\n"
- f"Format: [bold]GPTQ {bits}-bit[/]\n\n"
- f"Use with vLLM:\n"
- f" [bold]from vllm import LLM[/]\n"
- f" [bold]llm = LLM(model='{output_path}', quantization='gptq')[/]",
- title="[bold green]GPTQ Export Complete![/]",
- )
- )
- return output_path
-
-
def _auto_deploy_ollama(
output_path: Path, model_name: str, deploy_target: str, deploy_name: Optional[str]
):
diff --git a/src/soup_cli/commands/quantize.py b/src/soup_cli/commands/quantize.py
index 2de04e87..d86d882d 100644
--- a/src/soup_cli/commands/quantize.py
+++ b/src/soup_cli/commands/quantize.py
@@ -15,11 +15,9 @@ import typer
from rich.console import Console
from rich.markup import escape
-from soup_cli.config.deprecation import DEPRECATED_VALUE_REJECTION_VERSION
-
console = Console()
-VALID_FORMATS = frozenset({"gguf", "gptq", "awq", "onnx", "tensorrt"})
+VALID_FORMATS = frozenset({"gguf", "onnx", "tensorrt"})
def quantize(
@@ -30,16 +28,12 @@ def quantize(
to: str = typer.Option(
"gguf",
"--to",
- help=(
- "Target format: gguf | gptq | awq | onnx | tensorrt "
- "(awq and gptq are deprecated and will be removed in "
- f"Soup v{DEPRECATED_VALUE_REJECTION_VERSION})."
- ),
+ help="Target format: gguf | onnx | tensorrt.",
),
bits: int = typer.Option(
4,
"--bits",
- help="Quantization bits (1-16; respected by gguf/gptq/awq).",
+ help="Quantization bits (1-16; respected by gguf).",
),
output: str = typer.Option(
None,
@@ -68,8 +62,6 @@ def quantize(
parts = ["soup", "export", "--model", model_path, "--format", canonical]
if canonical == "gguf":
parts.extend(["--quant", f"q{bits}_K_M"])
- elif canonical in ("gptq", "awq"):
- parts.extend(["--bits", str(bits)])
if output:
parts.extend(["--output", output])
rendered = " ".join(shlex.quote(part) for part in parts)
@@ -78,9 +70,3 @@ def quantize(
console.print(
"[dim]Tip: `soup export --help` lists every advanced quantization flag.[/]"
)
- # #338 -- the command above still runs, but the format is on its way out.
- from soup_cli.commands.export import deprecated_format_notice
-
- notice = deprecated_format_notice(canonical)
- if notice is not None:
- console.print(f"[yellow]Deprecated:[/] {notice}")
diff --git a/src/soup_cli/mcp_server/registry.py b/src/soup_cli/mcp_server/registry.py
index 01de01f8..dac730b3 100644
--- a/src/soup_cli/mcp_server/registry.py
+++ b/src/soup_cli/mcp_server/registry.py
@@ -1191,7 +1191,7 @@ def _mutating_specs(
"type": "object",
"properties": {
"model": {"type": "string", "description": "Adapter/model path or id."},
- "format": {"type": "string", "description": "gguf / onnx / awq / gptq / ..."},
+ "format": {"type": "string", "description": "gguf / onnx / tensorrt / ..."},
"output": {"type": "string", "description": "Optional output path."},
},
"required": ["model", "format"],
diff --git a/tests/test_awq_gptq_export.py b/tests/test_awq_gptq_export.py
deleted file mode 100644
index ad7f4016..00000000
--- a/tests/test_awq_gptq_export.py
+++ /dev/null
@@ -1,1110 +0,0 @@
-"""Tests for AWQ and GPTQ export — config, validation, CLI."""
-
-import builtins
-import re
-import tempfile
-from pathlib import Path
-from unittest.mock import MagicMock
-from unittest.mock import patch as mock_patch
-
-import pytest
-from typer import Exit as ClickExit
-
-from soup_cli.commands.export import SUPPORTED_FORMATS
-
-# Rich colours per character on a colour-capable runner and WRAPS at the
-# terminal width, drawing box borders between the fragments of a wrapped help
-# line. Strip ANSI escapes and box-drawing characters, then collapse
-# whitespace, before any substring match on CLI output.
-_ANSI_RE = re.compile(r"\x1b\[[0-9;?]*[A-Za-z]")
-_BOX_RE = re.compile(r"[\u2500-\u257f]")
-
-
-def _plain(text: str) -> str:
- """ANSI- and border-stripped, whitespace-collapsed CLI output."""
- return " ".join(_BOX_RE.sub(" ", _ANSI_RE.sub("", text)).split())
-
-
-def _mock_import(awq_mock=None, gptq_mock=None):
- """Create a side_effect for builtins.__import__ that intercepts awq/auto_gptq."""
- real_import = builtins.__import__
-
- def custom_import(name, *args, **kwargs):
- if name == "awq" and awq_mock is not None:
- return awq_mock
- if name == "auto_gptq" and gptq_mock is not None:
- return gptq_mock
- return real_import(name, *args, **kwargs)
-
- return custom_import
-
-
-# ─── Format Support Tests ────────────────────────────────────────────────
-
-
-class TestExportFormatsExtended:
- """Test that AWQ and GPTQ formats are registered."""
-
- def test_awq_format_supported(self):
- assert "awq" in SUPPORTED_FORMATS
-
- def test_gptq_format_supported(self):
- assert "gptq" in SUPPORTED_FORMATS
-
- def test_format_count(self):
- """v0.53.1 — 7 prior formats + torchao + gguf-ud = 9."""
- assert len(SUPPORTED_FORMATS) == 9
-
- def test_all_formats_present(self):
- """v0.53.1 — adds torchao + gguf-ud to the v0.52.0 baseline."""
- expected = {
- "gguf", "onnx", "tensorrt", "awq", "gptq",
- "bitnet", "tq1_0",
- "torchao", "gguf-ud",
- }
- assert set(SUPPORTED_FORMATS) == expected
-
-
-# ─── AWQ Export CLI Tests ─────────────────────────────────────────────────
-
-
-class TestAwqExportCLI:
- """Test AWQ export via CLI."""
-
- def test_awq_export_missing_model_path(self, tmp_path):
- """soup export --format awq should fail if model path doesn't exist."""
- from typer.testing import CliRunner
-
- from soup_cli.cli import app
-
- runner = CliRunner()
- result = runner.invoke(
- app, ["export", "--model", str(tmp_path / "nonexistent"), "--format", "awq"]
- )
- assert result.exit_code != 0
-
- def test_awq_format_in_help(self):
- """Export help should mention awq format."""
- from typer.testing import CliRunner
-
- from soup_cli.cli import app
-
- runner = CliRunner()
- result = runner.invoke(app, ["export", "--help"])
- assert "awq" in result.output.lower()
-
- def test_gptq_format_in_help(self):
- """Export help should mention gptq format."""
- from typer.testing import CliRunner
-
- from soup_cli.cli import app
-
- runner = CliRunner()
- result = runner.invoke(app, ["export", "--help"])
- assert "gptq" in result.output.lower()
-
- def test_calibration_help_names_both_required_formats(self):
- """The shared option must not imply AWQ can omit calibration data.
-
- Asserted on the Click parameter rather than on rendered ``--help``:
- Rich wraps prose inside its box borders at any terminal width, so a
- substring match on the rendered text breaks on an unrelated width or
- help-string change.
- """
- import typer
-
- from soup_cli.cli import app
-
- cmd = typer.main.get_command(app).commands["export"]
- param = next(p for p in cmd.params if "--calibration-data" in p.opts)
-
- assert param.help == "Path to calibration JSONL. Required for AWQ and GPTQ."
-
-
-# ─── GPTQ Export CLI Tests ────────────────────────────────────────────────
-
-
-class TestGptqExportCLI:
- """Test GPTQ export via CLI."""
-
- def test_gptq_export_missing_model_path(self, tmp_path):
- """soup export --format gptq should fail if model path doesn't exist."""
- from typer.testing import CliRunner
-
- from soup_cli.cli import app
-
- runner = CliRunner()
- result = runner.invoke(
- app, ["export", "--model", str(tmp_path / "nonexistent"), "--format", "gptq"]
- )
- assert result.exit_code != 0
-
-
-# ─── AWQ Export Function Tests ────────────────────────────────────────────
-
-
-class TestAwqExportFunction:
- """Test _export_awq logic."""
-
- def test_export_awq_import_error(self, tmp_path, monkeypatch):
- """AWQ export should print friendly error when autoawq not installed."""
- from typer.testing import CliRunner
-
- from soup_cli.cli import app
-
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
- monkeypatch.chdir(tmp_path)
-
- runner = CliRunner()
- result = runner.invoke(
- app,
- [
- "export", "--model", str(model_dir), "--format", "awq",
- "--calibration-data", str(cal_file),
- ],
- )
- assert result.exit_code != 0
- assert "autoawq" in result.output.lower() or "not installed" in result.output.lower()
-
- def test_export_awq_no_calibration_data_raises(self, tmp_path, capsys):
- """AWQ must not delegate missing calibration data to AutoAWQ's implicit
- large dataset download, and must fail before importing or loading a model."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- real_import = builtins.__import__
-
- def reject_awq_import(name, *args, **kwargs):
- if name == "awq":
- pytest.fail("AWQ was imported before calibration validation")
- return real_import(name, *args, **kwargs)
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(builtins, "__import__", side_effect=reject_awq_import):
- with pytest.raises(ClickExit):
- export_mod._export_awq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=None,
- )
- output = _plain(capsys.readouterr().out)
- assert "AWQ export requires --calibration-data" in output
- assert "default calibration dataset" in output
-
- def test_export_awq_empty_calibration_file_raises(self, tmp_path):
- """An empty or malformed calibration file fails before model loading."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text("\nnot json\n", encoding="utf-8")
-
- real_import = builtins.__import__
-
- def reject_awq_import(name, *args, **kwargs):
- if name == "awq":
- pytest.fail("AWQ was imported before calibration validation")
- return real_import(name, *args, **kwargs)
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(builtins, "__import__", side_effect=reject_awq_import):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- with pytest.raises(ClickExit):
- export_mod._export_awq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
-
- @pytest.mark.parametrize(
- "payload",
- ["[]", "null", '"plain JSON string"', "42"],
- ids=["array", "null", "string", "number"],
- )
- def test_export_awq_rejects_non_object_json_rows(self, tmp_path, capsys, payload):
- """Valid JSON scalars and arrays are not valid JSONL calibration rows."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text(f"{payload}\n", encoding="utf-8")
-
- real_import = builtins.__import__
-
- def reject_awq_import(name, *args, **kwargs):
- if name == "awq":
- pytest.fail("AWQ was imported before calibration validation")
- return real_import(name, *args, **kwargs)
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(builtins, "__import__", side_effect=reject_awq_import):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- with pytest.raises(ClickExit):
- export_mod._export_awq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
-
- assert "No usable calibration samples" in capsys.readouterr().out
-
- def test_export_awq_rejects_non_utf8_calibration_before_import(self, tmp_path, capsys):
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_bytes(b'\xff{"text": "sample"}\n')
-
- real_import = builtins.__import__
-
- def reject_awq_import(name, *args, **kwargs):
- if name == "awq":
- pytest.fail("AWQ was imported before calibration validation")
- return real_import(name, *args, **kwargs)
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(builtins, "__import__", side_effect=reject_awq_import):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- with pytest.raises(ClickExit):
- export_mod._export_awq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
-
- # Rich's highlighter colours the "8" in "UTF-8" as a number.
- assert "not valid UTF-8" in _plain(capsys.readouterr().out)
-
- def test_export_awq_rejects_unreadable_calibration_before_import(self, tmp_path, capsys):
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_dir = tmp_path / "calibration-directory"
- cal_dir.mkdir()
-
- real_import = builtins.__import__
-
- def reject_awq_import(name, *args, **kwargs):
- if name == "awq":
- pytest.fail("AWQ was imported before calibration validation")
- return real_import(name, *args, **kwargs)
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(builtins, "__import__", side_effect=reject_awq_import):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_dir,
- ):
- with pytest.raises(ClickExit):
- export_mod._export_awq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(cal_dir),
- )
-
- assert "Could not read calibration data" in capsys.readouterr().out
-
- def test_export_awq_calls_quantize(self, tmp_path):
- """AWQ export should pass the explicit calibration rows to AutoAWQ."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
-
- mock_model = MagicMock()
- mock_awq_class = MagicMock()
- mock_awq_class.from_pretrained = MagicMock(return_value=mock_model)
- mock_tokenizer = MagicMock()
-
- awq_mod = MagicMock()
- awq_mod.AutoAWQForCausalLM = mock_awq_class
-
- import soup_cli.commands.export as export_mod
-
- out_path = tmp_path / "out"
- with mock_patch.object(builtins, "__import__", side_effect=_mock_import(awq_mock=awq_mod)):
- with mock_patch(
- "transformers.AutoTokenizer.from_pretrained",
- return_value=mock_tokenizer,
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_output_path",
- return_value=out_path,
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- export_mod._export_awq(
- model_dir, str(out_path), None,
- bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
- mock_awq_class.from_pretrained.assert_called_once()
- mock_model.quantize.assert_called_once_with(
- mock_tokenizer,
- quant_config={"zero_point": True, "q_group_size": 128, "w_bit": 4},
- calib_data=["sample"],
- )
- mock_model.save_quantized.assert_called_once()
-
- def test_export_awq_default_output_path(self, tmp_path):
- """Default AWQ output path should be model_name + _awq suffix."""
- model_dir = tmp_path / "my_model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
-
- mock_model = MagicMock()
- mock_awq_class = MagicMock()
- mock_awq_class.from_pretrained = MagicMock(return_value=mock_model)
- mock_tokenizer = MagicMock()
-
- awq_mod = MagicMock()
- awq_mod.AutoAWQForCausalLM = mock_awq_class
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(builtins, "__import__", side_effect=_mock_import(awq_mock=awq_mod)):
- with mock_patch(
- "transformers.AutoTokenizer.from_pretrained",
- return_value=mock_tokenizer,
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- export_mod._export_awq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
- save_call = mock_model.save_quantized.call_args
- output_path = save_call[0][0]
- assert "my_model_awq" in output_path
-
- def test_export_awq_with_calibration_data(self, tmp_path):
- """AWQ export with calibration data should load and pass it."""
- import json
-
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- cal_file = tmp_path / "cal.jsonl"
- lines = [json.dumps({"text": f"sample {i}"}) for i in range(10)]
- cal_file.write_text("\n".join(lines), encoding="utf-8")
-
- mock_model = MagicMock()
- mock_awq_class = MagicMock()
- mock_awq_class.from_pretrained = MagicMock(return_value=mock_model)
- mock_tokenizer = MagicMock()
-
- awq_mod = MagicMock()
- awq_mod.AutoAWQForCausalLM = mock_awq_class
-
- import soup_cli.commands.export as export_mod
-
- out_path = tmp_path / "out"
- with mock_patch.object(builtins, "__import__", side_effect=_mock_import(awq_mock=awq_mod)):
- with mock_patch(
- "transformers.AutoTokenizer.from_pretrained",
- return_value=mock_tokenizer,
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_output_path",
- return_value=out_path,
- ):
- export_mod._export_awq(
- model_dir, str(out_path), None,
- bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
- quant_call = mock_model.quantize.call_args
- assert quant_call is not None
- assert "calib_data" in quant_call.kwargs
-
- def test_export_awq_invalid_bits(self, tmp_path):
- """Invalid bits value should raise ClickExit."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- import soup_cli.commands.export as export_mod
-
- with pytest.raises(ClickExit):
- export_mod._export_awq(
- model_dir, None, None, bits=3, group_size=128,
- calibration_data=None,
- )
-
- def test_export_awq_calibration_path_traversal(self, tmp_path):
- """Calibration data path should be confined to cwd."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- import soup_cli.commands.export as export_mod
-
- with pytest.raises(ClickExit):
- export_mod._export_awq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(Path("C:/Windows/System32/drivers/etc/hosts")),
- )
-
-
-# ─── GPTQ Export Function Tests ───────────────────────────────────────────
-
-
-class TestGptqExportFunction:
- """Test _export_gptq logic."""
-
- def test_export_gptq_import_error(self, tmp_path, monkeypatch):
- """GPTQ export should print friendly error when auto-gptq not installed."""
- from typer.testing import CliRunner
-
- from soup_cli.cli import app
-
- monkeypatch.chdir(tmp_path)
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
-
- real_import = builtins.__import__
-
- def custom_import(name, *args, **kwargs):
- if name == "auto_gptq":
- raise ModuleNotFoundError("No module named 'auto_gptq'", name="auto_gptq")
- return real_import(name, *args, **kwargs)
-
- monkeypatch.setattr(builtins, "__import__", custom_import)
-
- runner = CliRunner()
- result = runner.invoke(
- app,
- [
- "export", "--model", str(model_dir), "--format", "gptq",
- "--calibration-data", str(cal_file),
- ],
- )
- assert result.exit_code != 0
- assert "auto-gptq" in result.output.lower() or "not installed" in result.output.lower()
-
- def test_export_gptq_no_calibration_data_raises(self, tmp_path, capsys):
- """GPTQ has no built-in fallback dataset: no --calibration-data must raise,
- never fall through to model.quantize(tokenizer) (auto-gptq expects examples,
- not a tokenizer, and rejects one with 'object is not iterable')."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- mock_gptq_class = MagicMock()
- gptq_mod = MagicMock()
- gptq_mod.AutoGPTQForCausalLM = mock_gptq_class
- gptq_mod.BaseQuantizeConfig = MagicMock
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(
- builtins, "__import__", side_effect=_mock_import(gptq_mock=gptq_mod)
- ):
- with pytest.raises(ClickExit):
- export_mod._export_gptq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=None,
- )
- mock_gptq_class.from_pretrained.assert_not_called()
- assert "--calibration-data" in capsys.readouterr().out
-
- def test_export_gptq_empty_calibration_file_raises(self, tmp_path):
- """A --calibration-data file that yields zero usable samples (empty,
- or only malformed JSON lines) must raise the same way as no file at
- all, before the model is ever loaded."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text("\n\nnot json\n", encoding="utf-8")
-
- mock_gptq_class = MagicMock()
- gptq_mod = MagicMock()
- gptq_mod.AutoGPTQForCausalLM = mock_gptq_class
- gptq_mod.BaseQuantizeConfig = MagicMock
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(
- builtins, "__import__", side_effect=_mock_import(gptq_mock=gptq_mod)
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- with pytest.raises(ClickExit):
- export_mod._export_gptq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
- mock_gptq_class.from_pretrained.assert_not_called()
-
- def test_export_gptq_calls_quantize(self, tmp_path):
- """GPTQ export should call AutoGPTQForCausalLM methods."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
-
- mock_model = MagicMock()
- mock_gptq_class = MagicMock()
- mock_gptq_class.from_pretrained = MagicMock(return_value=mock_model)
- mock_tokenizer = MagicMock()
-
- gptq_mod = MagicMock()
- gptq_mod.AutoGPTQForCausalLM = mock_gptq_class
- gptq_mod.BaseQuantizeConfig = MagicMock
-
- import soup_cli.commands.export as export_mod
-
- out_path = tmp_path / "out"
- out_path.mkdir()
- with mock_patch.object(
- builtins, "__import__", side_effect=_mock_import(gptq_mock=gptq_mod)
- ):
- with mock_patch(
- "transformers.AutoTokenizer.from_pretrained",
- return_value=mock_tokenizer,
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_output_path",
- return_value=out_path,
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- export_mod._export_gptq(
- model_dir, str(out_path), None,
- bits=4, group_size=128, calibration_data=str(cal_file),
- )
- mock_gptq_class.from_pretrained.assert_called_once()
- mock_model.quantize.assert_called_once()
- mock_model.save_quantized.assert_called_once()
-
- def test_export_gptq_default_output_path(self, tmp_path):
- """Default GPTQ output path should be model_name + _gptq suffix."""
- model_dir = tmp_path / "my_model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
-
- mock_model = MagicMock()
- mock_gptq_class = MagicMock()
- mock_gptq_class.from_pretrained = MagicMock(return_value=mock_model)
- mock_tokenizer = MagicMock()
-
- gptq_mod = MagicMock()
- gptq_mod.AutoGPTQForCausalLM = mock_gptq_class
- gptq_mod.BaseQuantizeConfig = MagicMock
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(
- builtins, "__import__", side_effect=_mock_import(gptq_mock=gptq_mod)
- ):
- with mock_patch(
- "transformers.AutoTokenizer.from_pretrained",
- return_value=mock_tokenizer,
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- export_mod._export_gptq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
- save_call = mock_model.save_quantized.call_args
- output_path = save_call[0][0]
- assert "my_model_gptq" in output_path
-
- def test_export_gptq_invalid_bits(self, tmp_path):
- """Invalid bits value should raise ClickExit."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- import soup_cli.commands.export as export_mod
-
- with pytest.raises(ClickExit):
- export_mod._export_gptq(
- model_dir, None, None, bits=3, group_size=128,
- calibration_data=None,
- )
-
- def test_export_gptq_calibration_path_traversal(self, tmp_path):
- """Calibration data path should be confined to cwd."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- import soup_cli.commands.export as export_mod
-
- with pytest.raises(ClickExit):
- export_mod._export_gptq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(Path("C:/Windows/System32/drivers/etc/hosts")),
- )
-
-
-# ─── GPTQ Shard Name Regression (issue #338, P2) ─────────────────────────
-
-
-class TestGptqStandardShardName:
- """auto_gptq's save_quantized names its shard gptq_model-<bits>bit-<group>g.safetensors,
- which AutoModelForCausalLM.from_pretrained does not look for."""
-
- def test_standardize_renames_shard_to_model_safetensors(self, tmp_path):
- from soup_cli.commands.export import _standardize_gptq_shard_name
-
- out = tmp_path / "out"
- out.mkdir()
- (out / "gptq_model-4bit-128g.safetensors").write_bytes(b"weights")
-
- _standardize_gptq_shard_name(out, bits=4, group_size=128)
-
- assert (out / "model.safetensors").exists()
- assert not (out / "gptq_model-4bit-128g.safetensors").exists()
-
- def test_standardize_does_not_overwrite_existing_standard_name(self, tmp_path):
- from soup_cli.commands.export import _standardize_gptq_shard_name
-
- out = tmp_path / "out"
- out.mkdir()
- (out / "model.safetensors").write_bytes(b"already-standard")
- (out / "gptq_model-4bit-128g.safetensors").write_bytes(b"weights")
-
- _standardize_gptq_shard_name(out, bits=4, group_size=128)
-
- assert (out / "model.safetensors").read_bytes() == b"already-standard"
-
- def test_standardize_no_op_when_shard_absent(self, tmp_path):
- from soup_cli.commands.export import _standardize_gptq_shard_name
-
- out = tmp_path / "out"
- out.mkdir()
-
- _standardize_gptq_shard_name(out, bits=4, group_size=128)
-
- assert not (out / "model.safetensors").exists()
-
- def test_export_gptq_end_to_end_leaves_standard_name(self, tmp_path):
- """Full _export_gptq run: save_quantized's mock writes the real auto_gptq
- shard name, and the standard model.safetensors must exist afterward."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
- out_path = tmp_path / "out"
- out_path.mkdir()
-
- def fake_save_quantized(path):
- Path(path, "gptq_model-4bit-128g.safetensors").write_bytes(b"weights")
-
- mock_model = MagicMock()
- mock_model.save_quantized.side_effect = fake_save_quantized
- mock_gptq_class = MagicMock()
- mock_gptq_class.from_pretrained = MagicMock(return_value=mock_model)
- mock_tokenizer = MagicMock()
-
- gptq_mod = MagicMock()
- gptq_mod.AutoGPTQForCausalLM = mock_gptq_class
- gptq_mod.BaseQuantizeConfig = MagicMock
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(
- builtins, "__import__", side_effect=_mock_import(gptq_mock=gptq_mod)
- ):
- with mock_patch(
- "transformers.AutoTokenizer.from_pretrained",
- return_value=mock_tokenizer,
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_output_path",
- return_value=out_path,
- ):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- export_mod._export_gptq(
- model_dir, str(out_path), None,
- bits=4, group_size=128, calibration_data=str(cal_file),
- )
-
- assert (out_path / "model.safetensors").exists()
- assert not (out_path / "gptq_model-4bit-128g.safetensors").exists()
-
-
-# ─── Import-Error Masking Regression (issue #338, ImportError swallowing) ──
-
-
-class TestAwqGptqImportErrorSurfacesRealCause:
- """except ImportError must not print a fixed 'not installed' string when the
- package itself imports but a transitive import inside it fails for an
- unrelated reason, which used to discard the real exception."""
-
- def test_gptq_transitive_import_failure_message(self, tmp_path, capsys):
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
-
- real_import = builtins.__import__
-
- def custom_import(name, *args, **kwargs):
- if name == "auto_gptq":
- raise ImportError("libfoo.so: cannot open shared object file", name="libfoo")
- return real_import(name, *args, **kwargs)
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(builtins, "__import__", side_effect=custom_import):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- with pytest.raises(ClickExit):
- export_mod._export_gptq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
- output = capsys.readouterr().out
- assert "libfoo.so" in output
- assert "not installed" not in output.lower()
-
- def test_awq_transitive_import_failure_message(self, tmp_path, capsys):
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
-
- real_import = builtins.__import__
-
- def custom_import(name, *args, **kwargs):
- if name == "awq":
- raise ImportError("libbar.so: cannot open shared object file", name="libbar")
- return real_import(name, *args, **kwargs)
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(builtins, "__import__", side_effect=custom_import):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- with pytest.raises(ClickExit):
- export_mod._export_awq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
- output = capsys.readouterr().out
- assert "libbar.so" in output
- assert "not installed" not in output.lower()
-
- def test_gptq_real_missing_package_still_shows_install_hint(self, tmp_path, capsys):
- """When auto-gptq is genuinely absent (exc.name == 'auto_gptq'), the
- friendly install hint must still show; this must not regress."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
-
- real_import = builtins.__import__
-
- def custom_import(name, *args, **kwargs):
- if name == "auto_gptq":
- raise ModuleNotFoundError("No module named 'auto_gptq'", name="auto_gptq")
- return real_import(name, *args, **kwargs)
-
- import soup_cli.commands.export as export_mod
-
- with mock_patch.object(builtins, "__import__", side_effect=custom_import):
- with mock_patch(
- "soup_cli.commands.export._validate_calibration_path",
- return_value=cal_file,
- ):
- with pytest.raises(ClickExit):
- export_mod._export_gptq(
- model_dir, None, None, bits=4, group_size=128,
- calibration_data=str(cal_file),
- )
- output = capsys.readouterr().out
- assert "not installed" in output.lower()
-
-
-# ─── CLI Integration Tests ───────────────────────────────────────────────
-
-
-class TestAwqGptqCLIArgs:
- """Test new CLI arguments are declared in export function signature."""
-
- def test_bits_flag_exists(self):
- """Export function should have a --bits parameter."""
- import inspect
-
- from soup_cli.commands.export import export
-
- sig = inspect.signature(export)
- assert "bits" in sig.parameters
-
- def test_group_size_flag_exists(self):
- """Export function should have a --group-size parameter."""
- import inspect
-
- from soup_cli.commands.export import export
-
- sig = inspect.signature(export)
- assert "group_size" in sig.parameters
-
- def test_calibration_data_flag_exists(self):
- """Export function should have a --calibration-data parameter."""
- import inspect
-
- from soup_cli.commands.export import export
-
- sig = inspect.signature(export)
- assert "calibration_data" in sig.parameters
-
- def test_calibration_samples_flag_exists(self):
- """Export function should have a --calibration-samples parameter."""
- import inspect
-
- from soup_cli.commands.export import export
-
- sig = inspect.signature(export)
- assert "calibration_samples" in sig.parameters
-
-
-# ─── Security Tests ──────────────────────────────────────────────────────
-
-
-class TestAwqGptqSecurity:
- """Security validation for AWQ/GPTQ export."""
-
- def test_calibration_data_path_traversal_awq(self, tmp_path):
- from typer.testing import CliRunner
-
- from soup_cli.cli import app
-
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- runner = CliRunner()
- result = runner.invoke(
- app,
- [
- "export", "--model", str(model_dir),
- "--format", "awq",
- "--calibration-data", "C:/Windows/System32/drivers/etc/hosts",
- ],
- )
- assert result.exit_code != 0
-
- def test_calibration_data_path_traversal_gptq(self, tmp_path):
- from typer.testing import CliRunner
-
- from soup_cli.cli import app
-
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- runner = CliRunner()
- result = runner.invoke(
- app,
- [
- "export", "--model", str(model_dir),
- "--format", "gptq",
- "--calibration-data", "C:/Windows/System32/drivers/etc/hosts",
- ],
- )
- assert result.exit_code != 0
-
-
-# ─── Calibration Data Loader Tests ────────────────────────────────────────
-
-
-class TestCalibrationLoader:
- """Test _load_calibration_texts function."""
-
- def test_load_text_field(self, tmp_path):
- import json
-
- from soup_cli.commands.export import _load_calibration_texts
-
- cal_file = tmp_path / "cal.jsonl"
- lines = [json.dumps({"text": f"line {i}"}) for i in range(5)]
- cal_file.write_text("\n".join(lines), encoding="utf-8")
-
- texts = _load_calibration_texts(cal_file, max_samples=10)
- assert len(texts) == 5
- assert texts[0] == "line 0"
-
- def test_load_max_samples(self, tmp_path):
- import json
-
- from soup_cli.commands.export import _load_calibration_texts
-
- cal_file = tmp_path / "cal.jsonl"
- lines = [json.dumps({"text": f"line {i}"}) for i in range(100)]
- cal_file.write_text("\n".join(lines), encoding="utf-8")
-
- texts = _load_calibration_texts(cal_file, max_samples=10)
- assert len(texts) == 10
-
- def test_load_none_returns_empty(self):
- from soup_cli.commands.export import _load_calibration_texts
-
- assert _load_calibration_texts(None) == []
-
- def test_load_skips_empty_lines(self, tmp_path):
- import json
-
- from soup_cli.commands.export import _load_calibration_texts
-
- cal_file = tmp_path / "cal.jsonl"
- content = json.dumps({"text": "hello"}) + "\n\n" + json.dumps({"text": "world"})
- cal_file.write_text(content, encoding="utf-8")
-
- texts = _load_calibration_texts(cal_file, max_samples=10)
- assert len(texts) == 2
-
- def test_load_concatenates_non_text_fields(self, tmp_path):
- import json
-
- from soup_cli.commands.export import _load_calibration_texts
-
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text(
- json.dumps({"prompt": "hello", "response": "world"}),
- encoding="utf-8",
- )
-
- texts = _load_calibration_texts(cal_file, max_samples=10)
- assert len(texts) == 1
- assert "hello" in texts[0]
- assert "world" in texts[0]
-
-
-# ─── Path Validation Tests ───────────────────────────────────────────────
-
-
-class TestCalibrationPathValidation:
- """Test _validate_calibration_path function."""
-
- def test_none_returns_none(self):
- from soup_cli.commands.export import _validate_calibration_path
-
- assert _validate_calibration_path(None) is None
-
- def test_nonexistent_file_raises(self):
- """Nonexistent calibration file should raise ClickExit."""
- from soup_cli.commands.export import _validate_calibration_path
-
- nope = Path.cwd() / "nonexistent_cal_data_test_xyzzy.jsonl"
- with pytest.raises(ClickExit):
- _validate_calibration_path(str(nope))
-
- def test_valid_path_returns_path(self):
- """Valid path under cwd should return resolved Path."""
- from soup_cli.commands.export import _validate_calibration_path
-
- cal_file = Path.cwd() / "test_cal_temp_xyzzy.jsonl"
- try:
- cal_file.write_text("{}", encoding="utf-8")
- result = _validate_calibration_path(str(cal_file))
- assert result is not None
- assert result.name == "test_cal_temp_xyzzy.jsonl"
- finally:
- if cal_file.exists():
- cal_file.unlink()
-
- def test_path_outside_cwd_raises(self):
- from soup_cli.commands.export import _validate_calibration_path
-
- with pytest.raises(ClickExit):
- _validate_calibration_path("C:/Windows/System32/drivers/etc/hosts")
-
-
-# ─── Output Path Validation Tests ────────────────────────────────────────
-
-
-class TestOutputPathValidation:
- """Test _validate_output_path function."""
-
- def test_none_returns_none(self):
- from soup_cli.commands.export import _validate_output_path
-
- assert _validate_output_path(None) is None
-
- def test_valid_path_returns_path(self):
- """Valid path under cwd should return resolved Path."""
- from soup_cli.commands.export import _validate_output_path
-
- out = Path.cwd() / "test_output_xyzzy"
- result = _validate_output_path(str(out))
- assert result is not None
- assert result.name == "test_output_xyzzy"
-
- def test_path_outside_cwd_raises(self):
- """Path outside cwd should raise ClickExit."""
- from soup_cli.commands.export import _validate_output_path
-
- # Use a path guaranteed to be outside cwd on all platforms
- outside = Path(tempfile.gettempdir()).resolve()
- cwd = Path.cwd().resolve()
- # Only test if temp dir is actually outside cwd
- try:
- outside.relative_to(cwd)
- pytest.skip("tempdir is under cwd, cannot test outside path")
- except ValueError:
- pass
- with pytest.raises(ClickExit):
- _validate_output_path(str(outside / "evil_output"))
-
- def test_path_traversal_rejected(self, tmp_path, monkeypatch):
- """Path outside cwd via relative traversal should be rejected."""
- from soup_cli.commands.export import _validate_output_path
-
- # Set cwd to a subdirectory so ../evil is outside
- sub = tmp_path / "deep" / "nested"
- sub.mkdir(parents=True)
- monkeypatch.chdir(sub)
- with pytest.raises(ClickExit):
- _validate_output_path(str(tmp_path / "evil"))
-
-
-# ─── Optional Dependency Tests ────────────────────────────────────────────
-
-
-class TestOptionalDeps:
- """Test that optional deps are declared in pyproject.toml."""
-
- def _read_pyproject(self):
- pyproject = Path(__file__).parent.parent / "pyproject.toml"
- return pyproject.read_text(encoding="utf-8")
-
- def test_awq_optional_dep(self):
- content = self._read_pyproject()
- assert "awq" in content
- assert "autoawq" in content
-
- def test_gptq_optional_dep(self):
- content = self._read_pyproject()
- assert "gptq" in content
- assert "auto-gptq" in content
diff --git a/tests/test_code_review_recurring.py b/tests/test_code_review_recurring.py
index 69d78e77..cd97a66c 100644
--- a/tests/test_code_review_recurring.py
+++ b/tests/test_code_review_recurring.py
@@ -30,7 +30,6 @@ def test_containment_holdouts_migrated_to_commonpath_helper():
"utils/ollama.py": "is_under_cwd",
"migrate/common.py": "is_under_cwd",
"commands/serve.py": "is_under(",
- "commands/export.py": "is_under_cwd",
"commands/generate.py": "is_under(",
"data/loader.py": "is_under(",
"eval/checkpoint_intelligence.py": "is_under(",
diff --git a/tests/test_export_removed_formats.py b/tests/test_export_removed_formats.py
new file mode 100644
index 00000000..7a26b765
--- /dev/null
+++ b/tests/test_export_removed_formats.py
@@ -0,0 +1,135 @@
+"""#1710 — Remove deprecated AWQ and GPTQ export and quantize formats.
+
+Follow-up to #338 and #1690. AWQ and GPTQ export formats and extras are removed
+one release after the deprecation notice. The CLI refuses both formats with a
+non-zero exit code and names the formats that stay: gguf, onnx, tensorrt, bitnet, tq1_0.
+"""
+
+from __future__ import annotations
+
+import pathlib
+import re
+
+import pytest
+from typer.testing import CliRunner
+
+import soup_cli.commands.export as export_mod
+import soup_cli.commands.quantize as quantize_mod
+from soup_cli.cli import app
+
+ROOT = pathlib.Path(__file__).resolve().parents[1]
+PYPROJECT = ROOT / "pyproject.toml"
+
+_ANSI_RE = re.compile(r"\x1b\[[0-9;?]*[A-Za-z]")
+_BOX_RE = re.compile(r"[\u2500-\u257f]")
+
+REMOVED = ("awq", "gptq")
+STAYING_EXPORT = ("gguf", "onnx", "tensorrt", "bitnet", "tq1_0")
+VALID_QUANTIZE = ("gguf", "onnx", "tensorrt")
+
+
+def _plain(text: str) -> str:
+ """ANSI- and border-stripped, whitespace-collapsed CLI output."""
+ return " ".join(_BOX_RE.sub(" ", _ANSI_RE.sub("", text)).split())
+
+
+class TestExportRefusesRemovedFormats:
+ @pytest.mark.parametrize("fmt", REMOVED)
+ def test_export_refuses_removed_format(self, fmt: str, tmp_path: pathlib.Path) -> None:
+ model_dir = tmp_path / "model"
+ model_dir.mkdir()
+
+ result = CliRunner().invoke(app, ["export", "--model", str(model_dir), "--format", fmt])
+
+ assert result.exit_code == 1, (result.output, repr(result.exception))
+ output = _plain(result.output)
+ assert f"unsupported format: {fmt}" in output.lower()
+ assert f"--format {fmt} was removed in v0.77." in output
+ for remaining in STAYING_EXPORT:
+ assert remaining in output
+
+ @pytest.mark.parametrize("fmt", REMOVED)
+ def test_export_refusal_does_not_mention_deprecation(
+ self, fmt: str, tmp_path: pathlib.Path
+ ) -> None:
+ model_dir = tmp_path / "model"
+ model_dir.mkdir()
+
+ result = CliRunner().invoke(app, ["export", "--model", str(model_dir), "--format", fmt])
+
+ output = _plain(result.output).lower()
+ assert "deprecated" not in output
+
+ def test_export_help_omits_removed_formats(self) -> None:
+ result = CliRunner().invoke(app, ["export", "--help"])
+ assert result.exit_code == 0, (result.output, repr(result.exception))
+ output = _plain(result.output).lower()
+ assert "awq and gptq are deprecated" not in output
+ assert "--format awq" not in output
+ assert "--format gptq" not in output
+ assert "--bits" not in output
+ assert "--group-size" not in output
+ assert "--calibration-samples" not in output
+
+
+class TestQuantizeRefusesRemovedFormats:
+ @pytest.mark.parametrize("fmt", REMOVED)
+ def test_quantize_refuses_removed_format(self, fmt: str) -> None:
+ result = CliRunner().invoke(app, ["quantize", "./out", "--to", fmt])
+
+ assert result.exit_code == 2
+ output = _plain(result.output)
+ assert "--to must be one of" in output
+ for valid in VALID_QUANTIZE:
+ assert valid in output
+ assert "deprecated" not in output.lower()
+
+ @pytest.mark.parametrize("fmt", REMOVED)
+ def test_quantize_refuses_uppercase_removed_format(self, fmt: str) -> None:
+ result = CliRunner().invoke(app, ["quantize", "./out", "--to", fmt.upper()])
+
+ assert result.exit_code == 2
+ output = _plain(result.output)
+ assert "--to must be one of" in output
+
+ def test_quantize_help_omits_removed_formats(self) -> None:
+ result = CliRunner().invoke(app, ["quantize", "--help"])
+ assert result.exit_code == 0
+ output = _plain(result.output).lower()
+ assert "awq and gptq are deprecated" not in output
+ assert "gptq | awq" not in output
+
+
+class TestCodebaseRemovals:
+ def test_supported_formats_excludes_awq_and_gptq(self) -> None:
+ for fmt in REMOVED:
+ assert fmt not in export_mod.SUPPORTED_FORMATS
+
+ def test_valid_formats_excludes_awq_and_gptq(self) -> None:
+ for fmt in REMOVED:
+ assert fmt not in quantize_mod.VALID_FORMATS
+
+ def test_deprecated_helpers_and_tables_removed(self) -> None:
+ assert not hasattr(export_mod, "DEPRECATED_FORMATS")
+ assert not hasattr(export_mod, "REPLACEMENT_FORMATS")
+ assert not hasattr(export_mod, "deprecated_format_notice")
+ assert not hasattr(export_mod, "_warn_deprecated_format")
+ assert not hasattr(export_mod, "_export_awq")
+ assert not hasattr(export_mod, "_export_gptq")
+ assert not hasattr(export_mod, "_standardize_gptq_shard_name")
+ assert not hasattr(export_mod, "_validate_output_path")
+ assert not hasattr(export_mod, "_validate_calibration_path")
+
+
+class TestPackagingRemovals:
+ def test_pyproject_excludes_awq_and_gptq_extras(self) -> None:
+ text = PYPROJECT.read_text(encoding="utf-8")
+ table = re.search(r"^\[project\.optional-dependencies\]\n(.*?)(?=^\[)", text, re.M | re.S)
+ assert table is not None
+ code = "\n".join(
+ line for line in table.group(1).splitlines() if not line.lstrip().startswith("#")
+ )
+ extras = re.findall(r"^([A-Za-z0-9_-]+)\s*=", code, re.M)
+ for fmt in REMOVED:
+ assert fmt not in extras
+ assert f"auto{fmt}" not in code
diff --git a/tests/test_issue1448_export_deploy_options.py b/tests/test_issue1448_export_deploy_options.py
index 47679745..d7f471ad 100644
--- a/tests/test_issue1448_export_deploy_options.py
+++ b/tests/test_issue1448_export_deploy_options.py
@@ -32,7 +32,7 @@ def test_torchao_in_valid_kinds():
assert "torchao" in _VALID_KINDS
[email protected]("fmt", ["onnx", "tensorrt", "awq", "gptq", "torchao"])
[email protected]("fmt", ["onnx", "tensorrt", "torchao"])
def test_export_refuses_deploy_for_unsupported_formats(tmp_path: Path, fmt: str):
"""Formats Ollama cannot load must refuse --deploy with an informative error."""
model_dir = tmp_path / "model"
@@ -103,8 +103,6 @@ def test_export_wires_deploy_for_deployable_formats(tmp_path: Path, fmt: str):
[
("onnx", "soup_cli.commands.export._export_onnx", "onnx"),
("tensorrt", "soup_cli.commands.export._export_tensorrt", "tensorrt"),
- ("awq", "soup_cli.commands.export._export_awq", "awq"),
- ("gptq", "soup_cli.commands.export._export_gptq", "gptq"),
("torchao", "soup_cli.commands.export._export_torchao_cli", "torchao"),
("gguf-ud", "soup_cli.commands.export._export_gguf_advanced", "gguf"),
("bitnet", "soup_cli.commands.export._export_bitnet_gguf", "gguf"),
diff --git a/tests/test_issue1448_real_helpers.py b/tests/test_issue1448_real_helpers.py
index c5c05b97..0471df59 100644
--- a/tests/test_issue1448_real_helpers.py
+++ b/tests/test_issue1448_real_helpers.py
@@ -35,31 +35,6 @@ def _module(name, **attrs):
return module
-def _quantizer(module_name, cls_name, saved_name):
- class _Model:
- def quantize(self, *args, **kwargs):
- return None
-
- def save_quantized(self, path):
- _write(Path(path) / saved_name, 3000)
-
- return {
- module_name: _module(
- module_name,
- BaseQuantizeConfig=lambda **kwargs: None,
- **{cls_name: types.SimpleNamespace(from_pretrained=lambda *a, **k: _Model())},
- )
- }
-
-
-def _gptq_modules():
- return _quantizer("auto_gptq", "AutoGPTQForCausalLM", "gptq_model-4bit-128g.safetensors")
-
-
-def _awq_modules():
- return _quantizer("awq", "AutoAWQForCausalLM", "model.safetensors")
-
-
def _onnx_modules():
def main_export(model_name_or_path, output, task):
_write(Path(output) / "model.onnx", 100)
@@ -106,10 +81,6 @@ CASES = {
[("onnx", "model.onnx"), ("onnx", "model.onnx_data")]),
"tensorrt": ([], _tensorrt_modules,
[("tensorrt", "rank0.engine")]),
- "awq": (["--calibration-data", "calib.jsonl"], _awq_modules,
- [("awq", "model.safetensors")]),
- "gptq": (["--calibration-data", "calib.jsonl"], _gptq_modules,
- [("gptq", "model.safetensors")]),
"torchao": (["--quant-config", "q.yaml"], dict,
[("torchao", "pytorch_model.bin")]),
"gguf-ud": (["--gguf-flavour", "Q4_0_4_4"], dict, [("gguf", "out.gguf")]),
diff --git a/tests/test_issue338_awq_gptq_deprecated.py b/tests/test_issue338_awq_gptq_deprecated.py
deleted file mode 100644
index 38ebdef2..00000000
--- a/tests/test_issue338_awq_gptq_deprecated.py
+++ /dev/null
@@ -1,197 +0,0 @@
-"""#338 -- AWQ and GPTQ export are deprecated, and the commands say so.
-
-Both upstream projects are archived and neither extra installs next to the
-train extra, so the two formats go away one release after this notice. Until
-then they keep working: ``soup export --format awq / gptq`` prints the notice
-first and then does what it did, and ``soup quantize --to awq / gptq`` prints
-it next to the command it hands the user.
-"""
-
-from __future__ import annotations
-
-import re
-
-import pytest
-from typer.testing import CliRunner
-
-import soup_cli.commands.export as export_mod
-from soup_cli.cli import app
-from soup_cli.commands.export import (
- DEPRECATED_FORMATS,
- REPLACEMENT_FORMATS,
- SUPPORTED_FORMATS,
- deprecated_format_notice,
-)
-from soup_cli.config.deprecation import DEPRECATED_VALUE_REJECTION_VERSION
-
-# Rich colours per character on a colour-capable runner and wraps at the
-# terminal width, drawing box borders between the fragments of a wrapped help
-# line. Strip ANSI escapes and box-drawing characters, then collapse
-# whitespace, before any substring match on CLI output.
-_ANSI_RE = re.compile(r"\x1b\[[0-9;?]*[A-Za-z]")
-_BOX_RE = re.compile(r"[\u2500-\u257f]")
-
-DEPRECATED = ("awq", "gptq")
-# The formats the decision on #338 names as the way out.
-REPLACEMENTS = ("gguf", "onnx", "tensorrt", "bitnet", "tq1_0")
-
-
-def _plain(text: str) -> str:
- """ANSI- and border-stripped, whitespace-collapsed CLI output."""
- return " ".join(_BOX_RE.sub(" ", _ANSI_RE.sub("", text)).split())
-
-
-class TestTheNotice:
- def test_exactly_awq_and_gptq_are_deprecated(self) -> None:
- assert sorted(DEPRECATED_FORMATS) == sorted(DEPRECATED)
-
- def test_the_deprecated_formats_still_run(self) -> None:
- """Deprecated is not removed: the formats stay selectable for one more release."""
- assert set(DEPRECATED) <= set(SUPPORTED_FORMATS)
-
- def test_the_replacements_are_the_ones_the_decision_names(self) -> None:
- assert tuple(REPLACEMENT_FORMATS) == REPLACEMENTS
-
- def test_every_replacement_is_a_format_that_stays(self) -> None:
- assert set(REPLACEMENT_FORMATS) <= set(SUPPORTED_FORMATS)
- assert not set(REPLACEMENT_FORMATS) & set(DEPRECATED_FORMATS)
-
- @pytest.mark.parametrize("fmt", DEPRECATED)
- def test_it_names_the_format_the_removal_and_every_replacement(self, fmt: str) -> None:
- notice = deprecated_format_notice(fmt)
- assert notice is not None
- assert f"--format {fmt} is deprecated" in notice
- assert f"will be removed in Soup v{DEPRECATED_VALUE_REJECTION_VERSION}:" in notice
- for replacement in REPLACEMENTS:
- assert re.search(rf"\b{re.escape(replacement)}\b", notice), (replacement, notice)
-
- @pytest.mark.parametrize("fmt", DEPRECATED)
- def test_it_does_not_send_the_user_to_the_other_deprecated_format(self, fmt: str) -> None:
- other = "gptq" if fmt == "awq" else "awq"
- assert other not in deprecated_format_notice(fmt).lower()
-
- @pytest.mark.parametrize("fmt", DEPRECATED)
- def test_it_says_why(self, fmt: str) -> None:
- notice = deprecated_format_notice(fmt)
- assert "archived upstream" in notice
- assert f"the {fmt} extra cannot be installed next to the train extra" in notice
-
- @pytest.mark.parametrize("fmt", sorted(set(SUPPORTED_FORMATS) - set(DEPRECATED)))
- def test_a_format_that_stays_has_no_notice(self, fmt: str) -> None:
- assert deprecated_format_notice(fmt) is None
-
- @pytest.mark.parametrize("fmt", ["", "AWQ", "awq ", "exl2", "nope"])
- def test_an_unknown_spelling_has_no_notice(self, fmt: str) -> None:
- """The command lower-cases nothing: ``--format AWQ`` is refused as unsupported."""
- assert deprecated_format_notice(fmt) is None
-
- def test_the_notice_carries_no_console_markup(self) -> None:
- """It is printed through Rich: a square bracket in it would be read as a style."""
- for fmt in DEPRECATED:
- assert "[" not in deprecated_format_notice(fmt)
-
-
-class TestExportPrintsIt:
- @pytest.mark.parametrize("fmt", DEPRECATED)
- def test_the_notice_comes_first_and_the_export_path_still_runs(self, fmt, tmp_path) -> None:
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- result = CliRunner().invoke(app, ["export", "--model", str(model_dir), "--format", fmt])
-
- output = _plain(result.output)
- assert f"Deprecated: --format {fmt} is deprecated" in output, (
- result.output,
- repr(result.exception),
- )
- # What the command did before, it still does: with no calibration data
- # it stops at the same requirement, with the same exit code.
- requirement = f"{fmt.upper()} export requires --calibration-data"
- assert requirement in output
- assert output.index("Deprecated:") < output.index(requirement)
- assert result.exit_code == 1
-
- @pytest.mark.parametrize("fmt", DEPRECATED)
- def test_the_notice_leads_when_the_extra_is_missing(self, fmt, tmp_path, monkeypatch) -> None:
- """Most users meet these formats through the "not installed" message."""
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- cal_file = tmp_path / "cal.jsonl"
- cal_file.write_text('{"text": "sample"}\n', encoding="utf-8")
- monkeypatch.chdir(tmp_path)
-
- result = CliRunner().invoke(
- app,
- [
- "export", "--model", str(model_dir), "--format", fmt,
- "--calibration-data", str(cal_file),
- ],
- )
-
- output = _plain(result.output)
- assert result.exit_code != 0
- assert output.startswith(f"Deprecated: --format {fmt} is deprecated"), result.output
-
- @pytest.mark.parametrize("fmt", DEPRECATED)
- def test_the_notice_is_printed_once(self, fmt, tmp_path) -> None:
- model_dir = tmp_path / "model"
- model_dir.mkdir()
-
- result = CliRunner().invoke(app, ["export", "--model", str(model_dir), "--format", fmt])
-
- assert _plain(result.output).count("is deprecated") == 1
-
- def test_a_format_that_stays_prints_no_notice(self, tmp_path, monkeypatch) -> None:
- model_dir = tmp_path / "model"
- model_dir.mkdir()
- calls = []
- monkeypatch.setattr(export_mod, "_export_onnx", lambda *args, **kwargs: calls.append(args))
-
- result = CliRunner().invoke(app, ["export", "--model", str(model_dir), "--format", "onnx"])
-
- assert result.exit_code == 0, (result.output, repr(result.exception))
- assert len(calls) == 1
- assert "deprecated" not in _plain(result.output).lower()
-
- def test_the_help_marks_both_formats(self) -> None:
- result = CliRunner().invoke(app, ["export", "--help"])
-
- assert result.exit_code == 0, (result.output, repr(result.exception))
- output = _plain(result.output)
- assert "awq and gptq are deprecated" in output
- # The release the help promises is the one the warnings promise.
- assert f"removed in Soup v{DEPRECATED_VALUE_REJECTION_VERSION}" in output, output
-
-
-class TestQuantizePrintsIt:
- @pytest.mark.parametrize("fmt", DEPRECATED)
- def test_it_still_prints_the_command_and_adds_the_notice(self, fmt: str) -> None:
- result = CliRunner().invoke(app, ["quantize", "./out", "--to", fmt])
-
- output = _plain(result.output)
- assert result.exit_code == 0, (result.output, repr(result.exception))
- assert f"soup export --model ./out --format {fmt} --bits 4" in output
- assert f"Deprecated: --format {fmt} is deprecated" in output
-
- @pytest.mark.parametrize("fmt", DEPRECATED)
- def test_an_upper_case_target_gets_the_notice_too(self, fmt: str) -> None:
- """``soup quantize`` lower-cases ``--to``, so the notice follows the canonical name."""
- result = CliRunner().invoke(app, ["quantize", "./out", "--to", fmt.upper()])
-
- assert result.exit_code == 0, (result.output, repr(result.exception))
- assert f"Deprecated: --format {fmt} is deprecated" in _plain(result.output)
-
- @pytest.mark.parametrize("fmt", ["gguf", "onnx", "tensorrt"])
- def test_a_target_that_stays_prints_no_notice(self, fmt: str) -> None:
- result = CliRunner().invoke(app, ["quantize", "./out", "--to", fmt])
-
- assert result.exit_code == 0, (result.output, repr(result.exception))
- assert "deprecated" not in _plain(result.output).lower()
-
- def test_the_help_marks_both_targets(self) -> None:
- result = CliRunner().invoke(app, ["quantize", "--help"])
-
- assert result.exit_code == 0, (result.output, repr(result.exception))
- output = _plain(result.output)
- assert "awq and gptq are deprecated" in output
- assert f"removed in Soup v{DEPRECATED_VALUE_REJECTION_VERSION}" in output, output
diff --git a/tests/test_onnx_tensorrt_export.py b/tests/test_onnx_tensorrt_export.py
index 6349cee5..86532819 100644
--- a/tests/test_onnx_tensorrt_export.py
+++ b/tests/test_onnx_tensorrt_export.py
@@ -24,8 +24,8 @@ class TestExportFormats:
assert "tensorrt" in SUPPORTED_FORMATS
def test_format_count(self):
- """v0.53.1 — 7 prior formats + torchao + gguf-ud = 9."""
- assert len(SUPPORTED_FORMATS) == 9
+ """v0.77 — 9 prior formats - awq - gptq = 7."""
+ assert len(SUPPORTED_FORMATS) == 7
assert "bitnet" in SUPPORTED_FORMATS
assert "tq1_0" in SUPPORTED_FORMATS
assert "torchao" in SUPPORTED_FORMATS
diff --git a/tests/test_v0404_part_a.py b/tests/test_v0404_part_a.py
index 362df77f..8a6c3436 100644
--- a/tests/test_v0404_part_a.py
+++ b/tests/test_v0404_part_a.py
@@ -226,7 +226,7 @@ class TestExportHelperSignatures:
assert params["trust_remote_code"].default is False
@pytest.mark.parametrize(
- "fn_name", ["_export_onnx", "_export_tensorrt", "_export_awq", "_export_gptq"],
+ "fn_name", ["_export_onnx", "_export_tensorrt"],
)
def test_export_helper_signature(self, fn_name: str):
import inspect