ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
unsloth/tests
History
JoshuaL3000 4322f936c2
 test: fast end-to-end GRPO fast_inference vLLM rollout test (#7136) 
* Add fast fast_inference GRPO smoke test for the vLLM LoRA rollout path

Covers the vLLM >= 0.25.0 LoRA collision path (unsloth#7283, fixed in
unsloth-zoo#919) with all seven attention and MLP projections as LoRA targets so
both fused families (qkv_proj, gate_up_proj) are exercised. Kept tiny: the
ungated unsloth/Qwen2.5-0.5B-Instruct, max_steps=1 (the collision triggers on the
first rollout), short prompts/completions, and enforce_eager=True to skip CUDA
graph capture. Runs in ~89s cold and ~37s on a warm torch.compile cache.

Wrapped as a pytest test that skips without CUDA and still runs as a script; a
length-based reward gives non-zero GRPO advantages; asserts the vLLM engine is
attached at load and still bound on the trainer. Heavy imports are deferred into
the test so CPU-only collection stays import-free.

Co-authored-by: JoshuaL3000 <joshua.jian.ern.liew@intel.com>

* Assert GRPO metrics and pin seed in fast_inference test

Switch to unsloth/Qwen3-0.6B, disable vLLM torch.compile
(compilation_config=0) and run 3 steps so the updated LoRA adapter is
re-synced into vLLM on every step, not just loaded once.

Pin GRPOConfig(seed=...), which TRL forwards to vLLM SamplingParams, so
the run is reproducible, and assert per-step metrics (loss, grad_norm,
completion length, reward, reward spread, kl) instead of only checking
that train() returned. Verified across seeds 42/123/2024/7.

* Correct the seed comment and drop the pytest return

GRPOConfig(seed=...) does not reach vLLM SamplingParams: TRL's
generation_kwargs carries no seed key. Reproducibility comes from the
Trainer's set_seed pinning the global RNG the colocated sampler draws
from, so describe that instead.

Returning a value from a test triggers PytestReturnNotNoneWarning, which
pytest intends to make an error; the value was unused.

---------

Co-authored-by: danielhanchen <unslothai@gmail.com>
2026-07-26 00:22:48 -07:00
..
fast_inference test: fast end-to-end GRPO fast_inference vLLM rollout test (#7136) 2026-07-26 00:22:48 -07:00
notebooks Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
python AMD: CI coverage for recent fixes, plus three wrong gfx ids (#7431) 2026-07-25 18:58:02 -05:00
qlora Formatting: ruff line-length 100, kwarg-spacing passes, drop blank after short local imports (#6079) 2026-06-08 04:24:13 -07:00
saving Replace standalone Studio wording with Unsloth (#7221) 2026-07-19 00:47:04 -07:00
security Harden desktop release token permissions (#7172) 2026-07-16 05:01:48 -07:00
sh Fix ROCm wheel-index unit test: extract the gfx-arch probe helpers get_torch_index_url now calls (#7399) 2026-07-24 02:12:16 -07:00
studio Studio: fail fast on out-of-disk instead of a doomed llama.cpp source build (#7420) 2026-07-26 00:11:38 -07:00
studio_setup_ps1 Make Visual Studio + CMake optional on Windows (prebuilt llama.cpp needs no build tools) (#6499) 2026-06-22 01:11:09 -07:00
utils Experimental: correct varlen sample packing for hybrid linear-attention models (#7249) 2026-07-20 00:57:02 -07:00
version_compat test(version-compat): keep GRPO fake-run logits finite on CPU (#7247) 2026-07-19 04:54:17 -07:00
vllm_compat Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
__init__.py Qwen 3, Bug Fixes (#2445) 2025-04-30 22:38:39 -07:00
_zoo_aggressive_cuda_spoof.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
_zoo_rocm_spoof.py AMD: CI coverage for recent fixes, plus three wrong gfx ids (#7431) 2026-07-25 18:58:02 -05:00
conftest.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
run_all.sh AMD: CI coverage for recent fixes, plus three wrong gfx ids (#7431) 2026-07-25 18:58:02 -05:00
test_attention_implementation.py fix(gpt-oss): prefer flex attention over sdpa (#5701) 2026-05-22 08:38:38 -07:00
test_attn_impl_honor_explicit.py Honor an explicit sdpa or flex_attention request when flash is disabled (#6847) 2026-07-06 05:45:45 -07:00
test_bad_mappings_redirect.py Fix BAD_MAPPINGS not redirecting the -unsloth-bnb-4bit dynamic quants (#6949) 2026-07-08 17:37:44 -03:00
test_callback_signature_drift.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
test_cli_export_unpacking.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
test_enforce_kwargs_spacing.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
test_fast_gemv_dispatch.py Fix fast inference crash on compressed-tensors FP8 models (#7025) 2026-07-09 04:10:59 -07:00
test_fast_generate_slow_guard.py fast_generate: clear error for vLLM-style inputs when fast_inference=False (#6786) 2026-07-03 08:16:32 -07:00
test_finetune_last_n_layers.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
test_fp8_device_context.py Guard FP8 Triton launches with tensor device context (#6888) 2026-07-08 18:32:51 -03:00
test_fp8_restore_dropped_scale.py Restore dropped FP8 weight_scale_inv tensors on load (#6978) 2026-07-09 06:44:44 -07:00
test_fp8_tiny_e8m0.py Handle odd shapes and non-float scales in FP8BlockQuantLinear (#6848) 2026-07-06 05:44:55 -07:00
test_fused_ce_not_return_dict_logits.py fix: use EMPTY_LOGITS on the fused-CE not-return_dict path (#2068) (#6482) 2026-06-23 01:26:55 -07:00
test_gemma4_chat_template.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
test_gemma_2b_mapper_key.py Fix duplicate unsloth/gemma-2b-bnb-4bit mapper key routing the base 4bit repo to the instruct model (#6891) 2026-07-08 17:40:39 -03:00
test_generate_kwarg_gate.py Fix gpt-oss offload_embedding and generate() kwargs, and guard offload_embedding on tied/vLLM models (#6774) 2026-07-01 22:39:00 -07:00
test_get_model_name.py Fix DeepScaleR-1.5B mapper entry pointing its 16bit repo at DeepHermes-3-8B (#7088) 2026-07-14 17:56:12 -03:00
test_gradient_checkpointing_restore.py Fix TrainingArguments silently disabling unsloth gradient checkpointing (#6829) 2026-07-03 16:35:02 +01:00
test_ignored_tokenizer_casing.py Match IGNORED_TOKENIZER_NAMES case-insensitively (#6620) 2026-06-23 19:40:50 -03:00
test_import_fixes_drift.py fix: keep LoRA reloads working with PEFT 0.19 (#6748) 2026-06-30 20:26:57 +01:00
test_installer_skip_autostart.py fix(install): fail non-tauri installer errors (#7123) 2026-07-15 04:37:03 -07:00
test_installer_unsloth_version.py Installer: report the installed Unsloth version (#7265) 2026-07-21 03:48:49 -07:00
test_loader_glob_skip.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
test_missing_torchvision_vlm.py Fix misleading 'only for image models' error for Qwen3-VL when torchvision is missing (#6525) 2026-06-23 01:28:09 -07:00
test_model_registry.py fix(registry): don't register deepseek models at import time (#7227) 2026-07-19 03:37:23 -07:00
test_moe_lora_targets.py Scope MoE expert LoRA detection to actual MLP projection targets (#6849) 2026-07-06 05:45:06 -07:00
test_multi_image_grpo_chunking.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
test_nvfp4_quant_load.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
test_offline_loading_helpers.py Fix offline checkpoint load/export: "tokenizer is weirdly not loaded" (#6554) 2026-06-25 23:16:53 -07:00
test_offload_embedding_hooks.py Fix gpt-oss offload_embedding and generate() kwargs, and guard offload_embedding on tied/vLLM models (#6774) 2026-07-01 22:39:00 -07:00
test_offload_tied_guard.py Fix gpt-oss offload_embedding and generate() kwargs, and guard offload_embedding on tied/vLLM models (#6774) 2026-07-01 22:39:00 -07:00
test_peft_tensor_parallel_compat.py fix: keep LoRA reloads working with PEFT 0.19 (#6748) 2026-06-30 20:26:57 +01:00
test_peft_weight_converter_compat.py Formatting: ruff line-length 100, kwarg-spacing passes, drop blank after short local imports (#6079) 2026-06-08 04:24:13 -07:00
test_prefetch_snapshot_scope.py Auto Xet to HTTP download fallback in from_pretrained; share Studio's fallback via unsloth_zoo (#6638) 2026-07-06 05:13:25 -07:00
test_pretrain_compile_reset.py Add regression tests for the stray-forward compile-cache reset (#6569) 2026-06-22 07:22:47 -07:00
test_public_api_surface.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
test_raw_text.py fix(dataprep): don't emit a degenerate chunk for empty text (#7183) 2026-07-23 00:56:51 -07:00
test_resolve_model_class.py Formatting: ruff line-length 100, kwarg-spacing passes, drop blank after short local imports (#6079) 2026-06-08 04:24:13 -07:00
test_studio_install_workspace_guard.py Studio: add local speech-to-text dictation engine (#7095) 2026-07-23 01:39:03 -07:00
test_studio_root_resilience.py Replace standalone Studio wording with Unsloth (#7221) 2026-07-19 00:47:04 -07:00
test_studio_shutdown_thread_wait.py Replace standalone Studio wording with Unsloth (#7221) 2026-07-19 00:47:04 -07:00
test_synthetic_chunk_data.py Fix SyntheticDataKit.chunk_data emitting chunks over max_tokens (#7073) 2026-07-12 05:06:11 -07:00
test_tool_mask_zoo_compat.py Formatting: ruff line-length 100, kwarg-spacing passes, drop blank after short local imports (#6079) 2026-06-08 04:24:13 -07:00
test_uma_safetensors_load.py Faster safetensors weight loading on unified-memory (integrated) GPUs (#5988) 2026-07-22 05:20:59 -07:00
test_uninitialized_position_ids.py Load DeepSeek-OCR and other VLMs that register AutoModel in auto_map (#6421) 2026-06-18 07:03:43 -07:00
test_video_path_validation.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00
test_vllm_broken_detection.py Fix fast_inference crash on ABI-broken vLLM: probe compiled extensions, not just import vllm (#6621) 2026-06-26 22:43:36 -07:00
test_windows_rocm_bnb_version.py Reduce and tighten comments and docstrings across the test suite (#6429) 2026-06-18 01:07:09 -07:00