diff --git a/python/sglang/srt/models/qwen3_vl_moe.py b/python/sglang/srt/models/qwen3_vl_moe.py index cf1cb3879b42..93746e5fcee8 100644 --- a/python/sglang/srt/models/qwen3_vl_moe.py +++ b/python/sglang/srt/models/qwen3_vl_moe.py @@ -179,9 +179,8 @@ def __init__( ): super().__init__(config, quant_config, prefix, language_model_cls) - # Only allow LoRA on attention projections within text layers for MoE. _lora_pattern_moe = re.compile( - r"^model\.layers\.(\d+)\.self_attn\.(?:qkv_proj|o_proj)$" + r"^(?:model\.layers\.(\d+)\.(?:self_attn\.(?:qkv_proj|o_proj)|mlp\.experts)|lm_head|model\.embed_tokens)$" ) def should_apply_lora(self, module_name: str) -> bool: diff --git a/test/manual/lora/test_lora_qwen3_vl.py b/test/manual/lora/test_lora_qwen3_vl.py deleted file mode 100644 index cef3649919a4..000000000000 --- a/test/manual/lora/test_lora_qwen3_vl.py +++ /dev/null @@ -1,233 +0,0 @@ -import random -import unittest -from typing import Sequence - -from sglang.srt.models.qwen3_vl import Qwen3VLForConditionalGeneration -from sglang.srt.models.qwen3_vl_moe import Qwen3VLMoeForConditionalGeneration -from sglang.test.lora_utils import ( - TORCH_DTYPES, - LoRAAdaptor, - LoRAModelCase, - ensure_reproducibility, -) -from sglang.test.runners import HFRunner, SRTRunner -from sglang.test.test_utils import CustomTestCase, calculate_rouge_l - - -class TestLoRAQwen3VLGating(CustomTestCase): - """Unit tests for should_apply_lora gating on Qwen3‑VL dense and MoE variants.""" - - def _assert_pattern( - self, pattern, positives: Sequence[str], negatives: Sequence[str] - ): - for name in positives: - self.assertTrue(bool(pattern.match(name)), f"Expected to match: {name}") - for name in negatives: - self.assertFalse(bool(pattern.match(name)), f"Should not match: {name}") - - def test_qwen3_vl_should_apply_lora_regex(self): - positives = ( - "model.layers.0.self_attn.qkv_proj", - "model.layers.1.self_attn.o_proj", - "model.layers.2.mlp.gate_up_proj", - "model.layers.3.mlp.down_proj", - ) - negatives = ( - "visual.blocks.0.attn.qkv_proj", - "model.layers.x.self_attn.qkv_proj", - "model.layers.0.attn.qkv_proj", - "model.layers.0.mlp.not_proj", - "model.layers.0.self_attn.q_proj", - ) - self._assert_pattern( - Qwen3VLForConditionalGeneration._lora_pattern, positives, negatives - ) - - def test_qwen3_vl_moe_should_apply_lora_regex(self): - positives = ( - "model.layers.0.self_attn.qkv_proj", - "model.layers.5.self_attn.o_proj", - ) - negatives = ( - "model.layers.0.mlp.gate_up_proj", - "model.layers.0.mlp.down_proj", - "visual.blocks.0.attn.qkv_proj", - "model.layers.x.self_attn.qkv_proj", - "model.layers.0.attn.qkv_proj", - ) - self._assert_pattern( - Qwen3VLMoeForConditionalGeneration._lora_pattern_moe, positives, negatives - ) - - -TEST_MULTIPLE_BATCH_PROMPTS = [ - """ - ### Instruction: - Tell me about llamas and alpacas - ### Response: - Llamas are large, long-necked animals with a woolly coat. They have two toes on each foot instead of three like other camelids (camels, dromedaries). Llamas live in the Andean mountains of South America where they graze on grasses and shrubs. Alpaca is another name for domesticated llama. The word "alpaca" comes from an Incan language meaning "golden fleece." Alpacas look very similar to llamas but are smaller than their wild relatives. Both species were used by ancient people as pack animals and for meat. Today both llamas and alpacas are raised primarily for their fiber which can be spun into yarn or knitted into clothing. - ### Question 2: - What do you know about llamas? - ### Answer: - """, - """ - ### Instruction: - Write a poem about the transformers Python library. - Mention the word "large language models" in that poem. - ### Response: - The Transformers are large language models, - They're used to make predictions on text. - """, - "AI is a field of computer science focused on", - "Computer science is the study of", - "Write a short story.", - "What are the main components of a computer?", -] - - -LORA_MODEL_VARIANTS = [ - ( - "Qwen3-VL", - LoRAModelCase( - base="Qwen/Qwen3-VL-4B-Instruct", - adaptors=[ - LoRAAdaptor( - name="mryufei/Qwen3-VL-4B-Instruct-trl-sft", - prefill_tolerance=3e-1, - ), - ], - max_loras_per_batch=1, - ), - ), - # TODO: Move 30B MoE to 2 GPU runner - # ( - # "Qwen3-VL-MoE", - # LoRAModelCase( - # base="Qwen/Qwen3-VL-30B-A3B-Instruct", - # adaptors=[ - # LoRAAdaptor( - # name="sosoai/qwen3_vl_30b_lora", - # prefill_tolerance=3e-1, - # ), - # ], - # max_loras_per_batch=1, - # ), - # ), -] - -LORA_MAX_NEW_TOKENS = 32 - - -def _run_lora_multiple_batch_on_model_cases( - model_cases: Sequence[LoRAModelCase], *, max_new_tokens: int, variant_label: str -): - for model_case in model_cases: - for torch_dtype in TORCH_DTYPES: - backend = "csgmv" - base_path = model_case.base - lora_adapter_paths = [adaptor.name for adaptor in model_case.adaptors] - - batches = [ - ( - [ - random.choice(TEST_MULTIPLE_BATCH_PROMPTS), - random.choice(TEST_MULTIPLE_BATCH_PROMPTS), - random.choice(TEST_MULTIPLE_BATCH_PROMPTS), - ], - [None, lora_adapter_paths[0], None], - ), - ( - [ - random.choice(TEST_MULTIPLE_BATCH_PROMPTS), - random.choice(TEST_MULTIPLE_BATCH_PROMPTS), - random.choice(TEST_MULTIPLE_BATCH_PROMPTS), - ], - [lora_adapter_paths[0], None, None], - ), - ( - [ - random.choice(TEST_MULTIPLE_BATCH_PROMPTS), - random.choice(TEST_MULTIPLE_BATCH_PROMPTS), - random.choice(TEST_MULTIPLE_BATCH_PROMPTS), - ], - [None, None, None], - ), - ] - - print( - f"\n=== {variant_label} LoRA parity on '{base_path}', backend={backend}, dtype={torch_dtype} ===" - ) - - ensure_reproducibility() - srt_runner = SRTRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - lora_paths=lora_adapter_paths, - max_loras_per_batch=model_case.max_loras_per_batch, - lora_backend=backend, - sleep_on_idle=True, - attention_backend="torch_native", - disable_radix_cache=True, - ) - - ensure_reproducibility() - hf_runner = HFRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - patch_model_do_sample_false=True, - ) - - with srt_runner, hf_runner: - for i, (prompts, lora_paths) in enumerate(batches): - print( - f"\n--- Running Batch {i + 1} --- prompts: {prompts}, lora_paths: {lora_paths}" - ) - - srt_outputs = srt_runner.batch_forward( - prompts, - max_new_tokens=max_new_tokens, - lora_paths=lora_paths, - ) - - hf_outputs = hf_runner.forward( - prompts, - max_new_tokens=max_new_tokens, - lora_paths=lora_paths, - ) - - print("SRT outputs:", [s for s in srt_outputs.output_strs]) - print("HF outputs:", [s for s in hf_outputs.output_strs]) - - for srt_out, hf_out in zip( - srt_outputs.output_strs, hf_outputs.output_strs - ): - srt_str = srt_out.strip() - hf_str = hf_out.strip() - rouge_tol = model_case.rouge_l_tolerance - rouge_score = calculate_rouge_l([srt_str], [hf_str])[0] - if rouge_score < rouge_tol: - raise AssertionError( - f"ROUGE-L score {rouge_score} below tolerance {rouge_tol} " - f"for base '{base_path}', adaptor '{lora_paths}', backend '{backend}', prompt: '{prompts}...'" - ) - - print(f"--- Batch {i + 1} Comparison Passed --- ") - - -class TestLoRAQwen3VLIntegration(CustomTestCase): - """Parity integration tests for Qwen3‑VL dense and MoE LoRA adapters.""" - - def test_ci_lora_models(self): - for label, model_case in LORA_MODEL_VARIANTS: - with self.subTest(variant=label): - _run_lora_multiple_batch_on_model_cases( - [model_case], - max_new_tokens=LORA_MAX_NEW_TOKENS, - variant_label=label, - ) - - -if __name__ == "__main__": - unittest.main() diff --git a/test/registered/lora/test_lora_qwen3_vl_30b_a3b_instruct_logprob_diff.py b/test/registered/lora/test_lora_qwen3_vl_30b_a3b_instruct_logprob_diff.py new file mode 100644 index 000000000000..67031048fa14 --- /dev/null +++ b/test/registered/lora/test_lora_qwen3_vl_30b_a3b_instruct_logprob_diff.py @@ -0,0 +1,151 @@ +# Copyright 2023-2025 SGLang Team +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== + +""" +Regression test for Qwen3-VL-30B-A3B-Instruct LoRA logprob accuracy. + +Compares SGLang LoRA logprobs against reference training logprobs from a +pre-computed dataset. The LoRA adapter and reference data are downloaded from: +https://huggingface.co/datasets/yushengsu/lora-diff-Qwen3-VL-30B-A3B-Instruct + +Usage: + python -m unittest test_lora_qwen3_vl_30b_a3b_instruct_logprob_diff +""" + +import multiprocessing as mp +import os +import unittest + +import torch +from huggingface_hub import snapshot_download + +import sglang as sgl +from sglang.test.ci.ci_register import register_cuda_ci +from sglang.test.test_utils import CustomTestCase + +register_cuda_ci( + est_time=300, + suite="stage-c-test-4-gpu-b200", +) + +BASE_MODEL = "Qwen/Qwen3-VL-30B-A3B-Instruct" +LORA_HF_REPO = "yushengsu/lora-diff-Qwen3-VL-30B-A3B-Instruct" +LORA_BACKEND = "triton" +MAX_LORA_RANK = 32 +TP_SIZE = 4 +DISABLE_CUDA_GRAPH = True +MOE_RUNNER_BACKEND = "triton" +EXPERTS_SHARED_OUTER_LORAS = True +PREFILL_ATTENTION_BACKEND = "fa4" +DECODE_ATTENTION_BACKEND = "fa4" + +KL_THRESHOLD = 1e-2 + + +def kl_v2(a, b): + a = torch.tensor(a) if not torch.is_tensor(a) else a + b = torch.tensor(b) if not torch.is_tensor(b) else b + return (((a - b) ** 2) * 0.5).mean().item() + + +def get_prompt_logprobs(engine, input_ids, lora_path): + out = engine.generate( + input_ids=input_ids, + sampling_params={"max_new_tokens": 0, "temperature": 0.0}, + return_logprob=True, + logprob_start_len=0, + lora_path=lora_path, + ) + return [logprob for logprob, _, _ in out["meta_info"]["input_token_logprobs"]][1:] + + +class TestLoRAQwen3VL_30B_A3B_Instruct_LogprobDiff(CustomTestCase): + + def test_lora_qwen3_vl_30b_a3b_instruct_logprob_accuracy(self): + adapter_path = snapshot_download( + LORA_HF_REPO, + repo_type="dataset", + ) + + engine = sgl.Engine( + model_path=BASE_MODEL, + tp_size=TP_SIZE, + enable_lora=True, + max_lora_rank=MAX_LORA_RANK, + lora_paths={"my_lora": adapter_path}, + lora_backend=LORA_BACKEND, + attention_backend="flashinfer", + disable_cuda_graph=DISABLE_CUDA_GRAPH, + moe_runner_backend=MOE_RUNNER_BACKEND, + experts_shared_outer_loras=EXPERTS_SHARED_OUTER_LORAS, + prefill_attention_backend=PREFILL_ATTENTION_BACKEND, + decode_attention_backend=DECODE_ATTENTION_BACKEND, + ) + + try: + cdata = torch.load( + os.path.join(adapter_path, "compare_sample_train_data.pt"), + weights_only=False, + ) + + base_logprobs = get_prompt_logprobs(engine, cdata["tokens"], lora_path=None) + logprobs = get_prompt_logprobs(engine, cdata["tokens"], lora_path="my_lora") + + base_t = torch.tensor(base_logprobs) + lora_t = torch.tensor(logprobs) + diff = (base_t - lora_t).abs() + print( + f"[VERIFY] base vs lora: mean_diff={diff.mean().item():.6f}, " + f"max_diff={diff.max().item():.6f}, " + f"identical={torch.equal(base_t, lora_t)}" + ) + + self.assertFalse( + torch.equal(base_t, lora_t), + "LoRA logprobs should differ from base model logprobs", + ) + + kl_sglang_trainer = kl_v2(cdata["training_logprobs"], logprobs) + kl_orig_trainer = kl_v2( + cdata["training_logprobs"], cdata["sampling_logprobs"] + ) + kl_sglang_orig = kl_v2(logprobs, cdata["sampling_logprobs"]) + + print(f"KL(orig_sampler, trainer) = {kl_orig_trainer:.6e}") + print(f"KL(sglang, trainer) = {kl_sglang_trainer:.6e}") + print(f"KL(sglang, orig_sampler) = {kl_sglang_orig:.6e}") + + self.assertLessEqual( + kl_sglang_trainer, + KL_THRESHOLD, + f"KL(sglang, trainer) = {kl_sglang_trainer:.6e} exceeds " + f"threshold {KL_THRESHOLD}", + ) + + finally: + engine.shutdown() + + +if __name__ == "__main__": + try: + mp.set_start_method("spawn") + except RuntimeError: + pass + + try: + unittest.main(warnings="ignore", verbosity=2) + finally: + if torch.cuda.is_available(): + torch.cuda.empty_cache() + torch.cuda.synchronize()