diff --git a/test/srt/models/lora/test_lora_backend.py b/test/srt/models/lora/test_lora_backend.py index 79d7dbebd..e159f8c6b 100644 --- a/test/srt/models/lora/test_lora_backend.py +++ b/test/srt/models/lora/test_lora_backend.py @@ -17,217 +17,38 @@ import os import unittest from typing import List -import torch -from utils import BACKENDS, TORCH_DTYPES, LoRAAdaptor, LoRAModelCase +from utils import ( + ALL_OTHER_LORA_MODELS, + BACKENDS, + CI_LORA_MODELS, + DEFAULT_PROMPTS, + TORCH_DTYPES, + LoRAModelCase, + run_batch_lora_test, +) -from sglang.test.runners import HFRunner, SRTRunner -from sglang.test.test_utils import CustomTestCase, calculate_rouge_l, is_in_ci - -CI_LORA_MODELS = [ - LoRAModelCase( - base="meta-llama/Llama-3.1-8B-Instruct", - adaptors=[ - LoRAAdaptor( - name="algoprog/fact-generation-llama-3.1-8b-instruct-lora", - ), - ], - max_loras_per_batch=1, - ), - LoRAModelCase( - base="meta-llama/Llama-3.1-8B-Instruct", - adaptors=[ - LoRAAdaptor( - name="Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16", - prefill_tolerance=1e-1, - ), - ], - max_loras_per_batch=1, - ), -] - -ALL_OTHER_LORA_MODELS = [ - LoRAModelCase( - base="meta-llama/Llama-2-7b-hf", - adaptors=[LoRAAdaptor(name="winddude/wizardLM-LlaMA-LoRA-7B")], - max_loras_per_batch=2, - ), -] - -PROMPTS = [ - "AI is a field of computer science focused on", - """ - ### 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: - """, -] +from sglang.test.test_utils import CustomTestCase, is_in_ci class TestLoRABackend(CustomTestCase): - def run_backend( - self, - prompt: str, - model_case: LoRAModelCase, - torch_dtype: torch.dtype, - max_new_tokens: int, - backend: str, - ): - """ - Run backend tests for a single prompt and model case. - """ - base_path = model_case.base - adaptor = model_case.adaptors[0] - print( - f"\n========== Testing backend '{backend}' for base '{base_path}' --- " - f"Prompt '{prompt[:50]}...' using adaptor '{adaptor.name}' ---" - ) - with SRTRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - tp_size=model_case.tp_size, - lora_paths=[adaptor.name for adaptor in model_case.adaptors], - max_loras_per_batch=model_case.max_loras_per_batch, - lora_backend=backend, - disable_cuda_graph=True, - disable_radix_cache=True, - mem_fraction_static=0.88, - disable_custom_all_reduce=False, - ) as srt_runner: - srt_outputs = srt_runner.forward( - [prompt], max_new_tokens=max_new_tokens, lora_paths=[adaptor.name] - ) - - with HFRunner( - base_path, torch_dtype=torch_dtype, model_type="generation" - ) as hf_runner: - hf_outputs = hf_runner.forward( - [prompt], max_new_tokens=max_new_tokens, lora_paths=[adaptor.name] - ) - - with SRTRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - tp_size=model_case.tp_size, - mem_fraction_static=0.88, - disable_custom_all_reduce=False, - ) as srt_runner: - srt_no_lora_outputs = srt_runner.forward( - [prompt], max_new_tokens=max_new_tokens - ) - - with HFRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - ) as hf_runner: - hf_no_lora_outputs = hf_runner.forward( - [prompt], max_new_tokens=max_new_tokens - ) - - # Use individual adapter tolerances if set, otherwise use model defaults - prefill_tol = ( - adaptor.prefill_tolerance - if adaptor.prefill_tolerance is not None - else model_case.prefill_tolerance - ) - decode_tol = ( - adaptor.decode_tolerance - if adaptor.decode_tolerance is not None - else model_case.decode_tolerance - ) - rouge_tol = ( - adaptor.rouge_l_tolerance - if adaptor.rouge_l_tolerance is not None - else model_case.rouge_l_tolerance - ) - - # Compare prefill stage logprobs (HF vs SRTRunner with LoRA) - hf_prefill = torch.tensor(hf_outputs.top_input_logprobs[0]) - srt_prefill = torch.tensor(srt_outputs.top_input_logprobs[0]) - max_prefill_diff = torch.max(torch.abs(hf_prefill - srt_prefill)) - print("Max prefill diff (HF vs SRT):", max_prefill_diff) - - # Compare decode stage logprobs - hf_decode = torch.tensor(hf_outputs.top_output_logprobs[0]) - srt_decode = torch.tensor(srt_outputs.top_output_logprobs[0]) - max_decode_diff = torch.max(torch.abs(hf_decode - srt_decode)) - print("Max decode diff (HF vs SRT):", max_decode_diff) - - srt_output_str = srt_outputs.output_strs[0].strip() - hf_output_str = hf_outputs.output_strs[0].strip() - rouge_score = calculate_rouge_l([srt_output_str], [hf_output_str])[0] - print("ROUGE-L score:", rouge_score) - print("SRT output:", srt_output_str) - print("HF output:", hf_output_str) - - # Additional: compare prefill outputs between base model (no LoRA) and LoRA model for reference - hf_no_lora_prefill = torch.tensor(hf_no_lora_outputs.top_input_logprobs[0]) - srt_no_lora_prefill = torch.tensor(srt_no_lora_outputs.top_input_logprobs[0]) - print( - "Max diff (SRT base vs SRT LoRA prefill):", - torch.max(torch.abs(srt_no_lora_prefill - srt_prefill)), - ) - print( - "Max diff (HF base vs HF LoRA prefill):", - torch.max(torch.abs(hf_no_lora_prefill - hf_prefill)), - ) - - if hf_prefill.shape[0] <= 100: - assert torch.all(torch.abs(hf_prefill - srt_prefill) < prefill_tol), ( - f"Prefill logprobs mismatch for base '{base_path}', adaptor '{adaptor.name}', " - f"backend '{backend}', prompt: '{prompt[:50]}...'" - ) - - if hf_decode.shape[0] <= 100: - assert torch.all(torch.abs(hf_decode - srt_decode) < decode_tol), ( - f"Decode logprobs mismatch for base '{base_path}', adaptor '{adaptor.name}', " - f"backend '{backend}', prompt: '{prompt[:50]}...'" - ) - - if rouge_score < rouge_tol: - - raise AssertionError( - f"ROUGE-L score {rouge_score} below tolerance {rouge_tol} " - f"for base '{base_path}', adaptor '{adaptor.name}', backend '{backend}', prompt: '{prompt[:50]}...'" - ) - - def run_backend_batch( - self, - prompts: List[str], - model_case: LoRAModelCase, - torch_dtype: torch.dtype, - max_new_tokens: int, - backend: str, - ): - # TODO: Implement batch processing version of run_backend - raise NotImplementedError( - "Batch processing version of run_backend is not implemented yet." - ) def _run_backend_on_model_cases(self, model_cases: List[LoRAModelCase]): for model_case in model_cases: # If skip_long_prompt is True, filter out prompts longer than 1000 characters prompts = ( - PROMPTS + DEFAULT_PROMPTS if not model_case.skip_long_prompt - else [p for p in PROMPTS if len(p) < 1000] + else [p for p in DEFAULT_PROMPTS if len(p) < 1000] ) for torch_dtype in TORCH_DTYPES: for backend in BACKENDS: - for prompt in prompts: - self.run_backend( - prompt, - model_case, - torch_dtype, - max_new_tokens=32, - backend=backend, - ) + run_batch_lora_test( + prompts, + model_case, + torch_dtype, + max_new_tokens=32, + backend=backend, + ) def test_ci_lora_models(self): self._run_backend_on_model_cases(CI_LORA_MODELS) diff --git a/test/srt/models/lora/test_lora_tp.py b/test/srt/models/lora/test_lora_tp.py index 5e34538b5..3d3f2ac0b 100644 --- a/test/srt/models/lora/test_lora_tp.py +++ b/test/srt/models/lora/test_lora_tp.py @@ -17,219 +17,40 @@ import os import unittest from typing import List -import torch -from utils import TORCH_DTYPES, LoRAAdaptor, LoRAModelCase +from utils import ( + ALL_OTHER_LORA_MODELS, + CI_LORA_MODELS, + DEFAULT_PROMPTS, + TORCH_DTYPES, + LoRAModelCase, + run_batch_lora_test, +) -from sglang.test.runners import HFRunner, SRTRunner -from sglang.test.test_utils import CustomTestCase, calculate_rouge_l, is_in_ci - -CI_LORA_MODELS = [ - LoRAModelCase( - base="meta-llama/Llama-3.1-8B-Instruct", - adaptors=[ - LoRAAdaptor( - name="algoprog/fact-generation-llama-3.1-8b-instruct-lora", - ), - ], - max_loras_per_batch=1, - ), -] - -ALL_OTHER_LORA_MODELS = [ - LoRAModelCase( - base="meta-llama/Llama-3.1-8B-Instruct", - adaptors=[ - LoRAAdaptor( - name="Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16", - prefill_tolerance=1e-1, - ), - ], - max_loras_per_batch=1, - ), - LoRAModelCase( - base="meta-llama/Llama-2-7b-hf", - adaptors=[LoRAAdaptor(name="winddude/wizardLM-LlaMA-LoRA-7B")], - max_loras_per_batch=2, - ), -] - -PROMPTS = [ - "AI is a field of computer science focused on", - """ - ### 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: - """, -] - -BACKEND = "triton" +from sglang.test.test_utils import CustomTestCase, is_in_ci class TestLoRATP(CustomTestCase): - def run_tp( - self, - prompt: str, - model_case: LoRAModelCase, - torch_dtype: torch.dtype, - max_new_tokens: int, - ): - """ - Run triton backend tests with specified TP size for a single prompt and model case. - """ - base_path = model_case.base - adaptor = model_case.adaptors[0] - tp_size = model_case.tp_size - print( - f"\n========== Testing triton backend with TP size {tp_size} for base '{base_path}' --- " - f"Prompt '{prompt[:50]}...' using adaptor '{adaptor.name}' ---" - ) - with SRTRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - tp_size=tp_size, - lora_paths=[adaptor.name for adaptor in model_case.adaptors], - max_loras_per_batch=model_case.max_loras_per_batch, - lora_backend=BACKEND, - disable_cuda_graph=True, - disable_radix_cache=True, - mem_fraction_static=0.88, - disable_custom_all_reduce=True, - ) as srt_runner: - srt_outputs = srt_runner.forward( - [prompt], max_new_tokens=max_new_tokens, lora_paths=[adaptor.name] - ) - - with HFRunner( - base_path, torch_dtype=torch_dtype, model_type="generation" - ) as hf_runner: - hf_outputs = hf_runner.forward( - [prompt], max_new_tokens=max_new_tokens, lora_paths=[adaptor.name] - ) - - with SRTRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - tp_size=tp_size, - mem_fraction_static=0.88, - disable_custom_all_reduce=True, - ) as srt_runner: - srt_no_lora_outputs = srt_runner.forward( - [prompt], max_new_tokens=max_new_tokens - ) - - with HFRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - ) as hf_runner: - hf_no_lora_outputs = hf_runner.forward( - [prompt], max_new_tokens=max_new_tokens - ) - - # Use individual adapter tolerances if set, otherwise use model defaults - prefill_tol = ( - adaptor.prefill_tolerance - if adaptor.prefill_tolerance is not None - else model_case.prefill_tolerance - ) - decode_tol = ( - adaptor.decode_tolerance - if adaptor.decode_tolerance is not None - else model_case.decode_tolerance - ) - rouge_tol = ( - adaptor.rouge_l_tolerance - if adaptor.rouge_l_tolerance is not None - else model_case.rouge_l_tolerance - ) - - # Compare prefill stage logprobs (HF vs SRTRunner with LoRA) - hf_prefill = torch.tensor(hf_outputs.top_input_logprobs[0]) - srt_prefill = torch.tensor(srt_outputs.top_input_logprobs[0]) - max_prefill_diff = torch.max(torch.abs(hf_prefill - srt_prefill)) - print("Max prefill diff (HF vs SRT):", max_prefill_diff) - - # Compare decode stage logprobs - hf_decode = torch.tensor(hf_outputs.top_output_logprobs[0]) - srt_decode = torch.tensor(srt_outputs.top_output_logprobs[0]) - max_decode_diff = torch.max(torch.abs(hf_decode - srt_decode)) - print("Max decode diff (HF vs SRT):", max_decode_diff) - - srt_output_str = srt_outputs.output_strs[0].strip() - hf_output_str = hf_outputs.output_strs[0].strip() - rouge_score = calculate_rouge_l([srt_output_str], [hf_output_str])[0] - print("ROUGE-L score:", rouge_score) - print("SRT output:", srt_output_str) - print("HF output:", hf_output_str) - - # Additional: compare prefill outputs between base model (no LoRA) and LoRA model for reference - hf_no_lora_prefill = torch.tensor(hf_no_lora_outputs.top_input_logprobs[0]) - srt_no_lora_prefill = torch.tensor(srt_no_lora_outputs.top_input_logprobs[0]) - print( - "Max diff (SRT base vs SRT LoRA prefill):", - torch.max(torch.abs(srt_no_lora_prefill - srt_prefill)), - ) - print( - "Max diff (HF base vs HF LoRA prefill):", - torch.max(torch.abs(hf_no_lora_prefill - hf_prefill)), - ) - - if hf_prefill.shape[0] <= 100: - assert torch.all(torch.abs(hf_prefill - srt_prefill) < prefill_tol), ( - f"Prefill logprobs mismatch for base '{base_path}', adaptor '{adaptor.name}', " - f"triton backend with TP {tp_size}, prompt: '{prompt[:50]}...'" - ) - - if hf_decode.shape[0] <= 100: - assert torch.all(torch.abs(hf_decode - srt_decode) < decode_tol), ( - f"Decode logprobs mismatch for base '{base_path}', adaptor '{adaptor.name}', " - f"triton backend with TP {tp_size}, prompt: '{prompt[:50]}...'" - ) - - if rouge_score < rouge_tol: - raise AssertionError( - f"ROUGE-L score {rouge_score} below tolerance {rouge_tol} " - f"for base '{base_path}', adaptor '{adaptor.name}', triton backend with TP {tp_size}, prompt: '{prompt[:50]}...'" - ) - - def run_tp_batch( - self, - prompts: List[str], - model_case: LoRAModelCase, - torch_dtype: torch.dtype, - max_new_tokens: int, - tp_size: int, - ): - # TODO: Implement batch processing version of run_tp - raise NotImplementedError( - "Batch processing version of run_tp is not implemented yet." - ) def _run_tp_on_model_cases(self, model_cases: List[LoRAModelCase]): tp_list = [2] # Define TP sizes to iterate over for model_case in model_cases: # If skip_long_prompt is True, filter out prompts longer than 1000 characters prompts = ( - PROMPTS + DEFAULT_PROMPTS if not model_case.skip_long_prompt - else [p for p in PROMPTS if len(p) < 1000] + else [p for p in DEFAULT_PROMPTS if len(p) < 1000] ) for tp_size in tp_list: model_case.tp_size = tp_size for torch_dtype in TORCH_DTYPES: - for prompt in prompts: - self.run_tp( - prompt, - model_case, - torch_dtype, - max_new_tokens=32, - ) + run_batch_lora_test( + prompts, + model_case, + torch_dtype, + max_new_tokens=32, + backend="triton", + test_tag=f"tp={tp_size}", + ) def test_ci_lora_models(self): self._run_tp_on_model_cases(CI_LORA_MODELS) diff --git a/test/srt/models/lora/test_multi_lora_backend.py b/test/srt/models/lora/test_multi_lora_backend.py index 68e78c2c9..c40a01d6f 100644 --- a/test/srt/models/lora/test_multi_lora_backend.py +++ b/test/srt/models/lora/test_multi_lora_backend.py @@ -13,16 +13,21 @@ # ============================================================================== import multiprocessing as mp +import os import unittest from typing import List -import torch -from utils import BACKENDS, TORCH_DTYPES, LoRAAdaptor, LoRAModelCase +from utils import ( + BACKENDS, + TORCH_DTYPES, + LoRAAdaptor, + LoRAModelCase, + run_batch_lora_test, +) -from sglang.test.runners import HFRunner, SRTRunner -from sglang.test.test_utils import CustomTestCase, calculate_rouge_l, is_in_ci +from sglang.test.test_utils import CustomTestCase, is_in_ci -MULTI_LORA_MODELS = [ +CI_MULTI_LORA_MODELS = [ # multi-rank case LoRAModelCase( base="meta-llama/Llama-2-7b-hf", @@ -38,6 +43,9 @@ MULTI_LORA_MODELS = [ ], max_loras_per_batch=2, ), +] + +ALL_OTHER_MULTI_LORA_MODELS = [ LoRAModelCase( base="meta-llama/Llama-3.1-8B-Instruct", adaptors=[ @@ -70,141 +78,8 @@ PROMPTS = [ class TestMultiLoRABackend(CustomTestCase): - def run_backend_batch( - self, - prompts: List[str], - model_case: LoRAModelCase, - torch_dtype: torch.dtype, - max_new_tokens: int, - backend: str, - ): - """ - The multi-LoRA backend test functionality is not supported yet. - This function uses all prompts at once and prints a message indicating that support is pending. - """ - base_path = model_case.base - adaptor_names = [adaptor.name for adaptor in model_case.adaptors] - print( - f"\n========== Testing multi-LoRA backend '{backend}' for base '{model_case.base}' --- " - f"Using prompts {[p[:50] for p in prompts]} with adaptors: {adaptor_names} ---" - ) - print( - "run_backend_batch: Multi-LoRA backend test functionality is pending support." - ) - with SRTRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - tp_size=model_case.tp_size, - lora_paths=[adaptor.name for adaptor in model_case.adaptors], - max_loras_per_batch=model_case.max_loras_per_batch, - lora_backend=backend, - disable_cuda_graph=True, - disable_radix_cache=True, - mem_fraction_static=0.88, - ) as srt_runner: - srt_outputs = srt_runner.forward( - prompts, max_new_tokens=max_new_tokens, lora_paths=adaptor_names - ) - with HFRunner( - base_path, torch_dtype=torch_dtype, model_type="generation" - ) as hf_runner: - hf_outputs = hf_runner.forward( - prompts, max_new_tokens=max_new_tokens, lora_paths=adaptor_names - ) - - with SRTRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - tp_size=model_case.tp_size, - mem_fraction_static=0.88, - ) as srt_runner: - srt_no_lora_outputs = srt_runner.forward( - prompts, max_new_tokens=max_new_tokens - ) - - with HFRunner( - base_path, - torch_dtype=torch_dtype, - model_type="generation", - ) as hf_runner: - hf_no_lora_outputs = hf_runner.forward( - prompts, max_new_tokens=max_new_tokens - ) - - # Compare prefill stage logprobs (HF vs SRTRunner with LoRA) - for i in range(len(prompts)): - adaptor = model_case.adaptors[i] - # Use individual adapter tolerances if set, otherwise use model defaults - prefill_tol = ( - adaptor.prefill_tolerance - if adaptor.prefill_tolerance is not None - else model_case.prefill_tolerance - ) - decode_tol = ( - adaptor.decode_tolerance - if adaptor.decode_tolerance is not None - else model_case.decode_tolerance - ) - rouge_tol = ( - adaptor.rouge_l_tolerance - if adaptor.rouge_l_tolerance is not None - else model_case.rouge_l_tolerance - ) - # Compare prefill stage logprobs (HF vs SRTRunner with LoRA) - hf_prefill = torch.tensor(hf_outputs.top_input_logprobs[i]) - srt_prefill = torch.tensor(srt_outputs.top_input_logprobs[i]) - max_prefill_diff = torch.max(torch.abs(hf_prefill - srt_prefill)) - print("Max prefill diff (HF vs SRT):", max_prefill_diff) - - # Compare decode stage logprobs - hf_decode = torch.tensor(hf_outputs.top_output_logprobs[i]) - srt_decode = torch.tensor(srt_outputs.top_output_logprobs[i]) - max_decode_diff = torch.max(torch.abs(hf_decode - srt_decode)) - print("Max decode diff (HF vs SRT):", max_decode_diff) - - srt_output_str = srt_outputs.output_strs[i].strip() - hf_output_str = hf_outputs.output_strs[i].strip() - rouge_score = calculate_rouge_l([srt_output_str], [hf_output_str])[0] - print("ROUGE-L score:", rouge_score) - print("SRT output:", srt_output_str) - print("HF output:", hf_output_str) - - # Additional: compare prefill outputs between base model (no LoRA) and LoRA model for reference - hf_no_lora_prefill = torch.tensor(hf_no_lora_outputs.top_input_logprobs[i]) - srt_no_lora_prefill = torch.tensor( - srt_no_lora_outputs.top_input_logprobs[i] - ) - print( - "Max diff (SRT base vs SRT LoRA prefill):", - torch.max(torch.abs(srt_no_lora_prefill - srt_prefill)), - ) - print( - "Max diff (HF base vs HF LoRA prefill):", - torch.max(torch.abs(hf_no_lora_prefill - hf_prefill)), - ) - - if hf_prefill.shape[0] <= 100: - assert torch.all(torch.abs(hf_prefill - srt_prefill) < prefill_tol), ( - f"Prefill logprobs mismatch for base '{base_path}', adaptor '{adaptor_names}', " - f"backend '{backend}', prompt: '{prompts[0][:50]}...'" - ) - - if hf_decode.shape[0] <= 100: - assert torch.all(torch.abs(hf_decode - srt_decode) < decode_tol), ( - f"Decode logprobs mismatch for base '{base_path}', adaptor '{adaptor_names}', " - f"backend '{backend}', prompt: '{prompts[0][:50]}...'" - ) - - if rouge_score < rouge_tol: - raise AssertionError( - f"ROUGE-L score {rouge_score} below tolerance {rouge_tol} " - f"for base '{base_path}', adaptor '{adaptor_names}', backend '{backend}', prompt: '{prompts[0][:50]}...'" - ) - - def _run_backend_on_model_cases(self, model_cases: List[LoRAModelCase]): + def _run_multi_lora_test_on_model_cases(self, model_cases: List[LoRAModelCase]): for model_case in model_cases: # If skip_long_prompt is True, filter out prompts longer than 1000 characters. batch_prompts = ( @@ -214,19 +89,30 @@ class TestMultiLoRABackend(CustomTestCase): ) for torch_dtype in TORCH_DTYPES: for backend in BACKENDS: - self.run_backend_batch( + run_batch_lora_test( batch_prompts, model_case, torch_dtype, max_new_tokens=32, backend=backend, + test_tag="multi-lora-backend", ) - def test_multi_lora_models(self): - # Optionally skip tests in CI environments. + def test_ci_lora_models(self): + self._run_multi_lora_test_on_model_cases(CI_MULTI_LORA_MODELS) + + def test_all_lora_models(self): if is_in_ci(): return - self._run_backend_on_model_cases(MULTI_LORA_MODELS) + + # Retain ONLY_RUN check here + filtered_models = [] + for model_case in ALL_OTHER_MULTI_LORA_MODELS: + if "ONLY_RUN" in os.environ and os.environ["ONLY_RUN"] != model_case.base: + continue + filtered_models.append(model_case) + + self._run_multi_lora_test_on_model_cases(filtered_models) if __name__ == "__main__": diff --git a/test/srt/models/lora/utils.py b/test/srt/models/lora/utils.py index 116389a26..af68709f4 100644 --- a/test/srt/models/lora/utils.py +++ b/test/srt/models/lora/utils.py @@ -17,6 +17,9 @@ from typing import List import torch +from sglang.test.runners import HFRunner, SRTRunner +from sglang.test.test_utils import calculate_rouge_l + @dataclasses.dataclass class LoRAAdaptor: @@ -47,3 +50,190 @@ class LoRAModelCase: TORCH_DTYPES = [torch.float16] BACKENDS = ["triton"] +DEFAULT_PROMPTS = [ + "AI is a field of computer science focused on", + """ + ### 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: + """, +] + +CI_LORA_MODELS = [ + LoRAModelCase( + base="meta-llama/Llama-3.1-8B-Instruct", + adaptors=[ + LoRAAdaptor( + name="algoprog/fact-generation-llama-3.1-8b-instruct-lora", + ), + ], + max_loras_per_batch=1, + ), +] + +ALL_OTHER_LORA_MODELS = [ + LoRAModelCase( + base="meta-llama/Llama-3.1-8B-Instruct", + adaptors=[ + LoRAAdaptor( + name="Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16", + prefill_tolerance=1e-1, + ), + ], + max_loras_per_batch=1, + ), + LoRAModelCase( + base="meta-llama/Llama-2-7b-hf", + adaptors=[LoRAAdaptor(name="winddude/wizardLM-LlaMA-LoRA-7B")], + max_loras_per_batch=2, + ), +] + + +def run_batch_lora_test( + prompts: List[str], + model_case: LoRAModelCase, + torch_dtype: torch.dtype, + max_new_tokens: int, + backend: str, + disable_cuda_graph: bool = True, + disable_radix_cache: bool = True, + mem_fraction_static: float = 0.88, + test_tag: str = "", +): + """ + Run Lora test for a forward batch. + For prompt0, prompt1, ..., promptN, + we will use adaptor0, adaptor1, ..., adaptorN included in model case, + We will then compare the outputs of HF and SRT with and without LoRA. + If number of prompts is larger than number of adaptors, + the prompt i will use adaptor i % (number of adaptors). + + Args: + prompts (List[str]): The batch of prompts to test. + model_case (LoRAModelCase): The model case to test. + torch_dtype (torch.dtype): The torch dtype to use. + max_new_tokens (int): The maximum number of new tokens to generate. + backend (str): The lora backend to use. + disable_cuda_graph (bool, optional): Whether to disable CUDA graph. Defaults to True. + disable_radix_cache (bool, optional): Whether to disable radix cache. Defaults to True. + mem_fraction_static (float, optional): The fraction of memory to use. Defaults to 0.88. + test_tag (str, optional): The tag to use for the test. Defaults to "". + """ + base_path = model_case.base + + # Create used adaptors for each prompt in batch + i, adaptors = 0, [] + for _ in range(len(prompts)): + adaptors.append(model_case.adaptors[i]) + i = (i + 1) % len(model_case.adaptors) + adaptor_names = [adaptor.name for adaptor in adaptors] + + print( + f"\n========== Testing {test_tag} on base '{model_case.base}' with backend={backend}, dtype={torch_dtype} --- " + f"Using prompts {[p[:50] for p in prompts]} with adaptors: {adaptor_names} ---" + ) + with SRTRunner( + base_path, + torch_dtype=torch_dtype, + model_type="generation", + tp_size=model_case.tp_size, + lora_paths=[adaptor.name for adaptor in model_case.adaptors], + max_loras_per_batch=model_case.max_loras_per_batch, + lora_backend=backend, + disable_cuda_graph=disable_cuda_graph, + disable_radix_cache=disable_radix_cache, + mem_fraction_static=mem_fraction_static, + ) as srt_runner: + srt_outputs = srt_runner.forward( + prompts, max_new_tokens=max_new_tokens, lora_paths=adaptor_names + ) + + with SRTRunner( + base_path, + torch_dtype=torch_dtype, + model_type="generation", + tp_size=model_case.tp_size, + mem_fraction_static=mem_fraction_static, + ) as srt_runner: + srt_no_lora_outputs = srt_runner.forward(prompts, max_new_tokens=max_new_tokens) + + with HFRunner( + base_path, torch_dtype=torch_dtype, model_type="generation" + ) as hf_runner: + hf_outputs = hf_runner.forward( + prompts, max_new_tokens=max_new_tokens, lora_paths=adaptor_names + ) + hf_no_lora_outputs = hf_runner.forward(prompts, max_new_tokens=max_new_tokens) + + # Compare prefill stage logprobs (HF vs SRTRunner with LoRA) + for i in range(len(prompts)): + adaptor = adaptors[i] + # Use individual adaptor tolerances if set, otherwise use model defaults + prefill_tol = ( + adaptor.prefill_tolerance + if adaptor.prefill_tolerance is not None + else model_case.prefill_tolerance + ) + decode_tol = ( + adaptor.decode_tolerance + if adaptor.decode_tolerance is not None + else model_case.decode_tolerance + ) + rouge_tol = ( + adaptor.rouge_l_tolerance + if adaptor.rouge_l_tolerance is not None + else model_case.rouge_l_tolerance + ) + # Compare prefill stage logprobs (HF vs SRTRunner with LoRA) + hf_prefill = torch.tensor(hf_outputs.top_input_logprobs[i]) + srt_prefill = torch.tensor(srt_outputs.top_input_logprobs[i]) + max_prefill_diff = torch.max(torch.abs(hf_prefill - srt_prefill)) + print("Max prefill diff (HF vs SRT):", max_prefill_diff) + + # Compare decode stage logprobs + hf_decode = torch.tensor(hf_outputs.top_output_logprobs[i]) + srt_decode = torch.tensor(srt_outputs.top_output_logprobs[i]) + max_decode_diff = torch.max(torch.abs(hf_decode - srt_decode)) + print("Max decode diff (HF vs SRT):", max_decode_diff) + + srt_output_str = srt_outputs.output_strs[i].strip() + hf_output_str = hf_outputs.output_strs[i].strip() + rouge_score = calculate_rouge_l([srt_output_str], [hf_output_str])[0] + print("ROUGE-L score:", rouge_score) + print("SRT output:", srt_output_str) + print("HF output:", hf_output_str) + + # Additional: compare prefill outputs between base model (no LoRA) and LoRA model for reference + hf_no_lora_prefill = torch.tensor(hf_no_lora_outputs.top_input_logprobs[i]) + srt_no_lora_prefill = torch.tensor(srt_no_lora_outputs.top_input_logprobs[i]) + print( + "Max diff (SRT base vs SRT LoRA prefill):", + torch.max(torch.abs(srt_no_lora_prefill - srt_prefill)), + ) + print( + "Max diff (HF base vs HF LoRA prefill):", + torch.max(torch.abs(hf_no_lora_prefill - hf_prefill)), + ) + + if hf_prefill.shape[0] <= 100: + assert torch.all(torch.abs(hf_prefill - srt_prefill) < prefill_tol), ( + f"Prefill logprobs mismatch for base '{base_path}', adaptor '{adaptor_names}', " + f"backend '{backend}', prompt: '{prompts[0][:50]}...'" + ) + + if hf_decode.shape[0] <= 100: + assert torch.all(torch.abs(hf_decode - srt_decode) < decode_tol), ( + f"Decode logprobs mismatch for base '{base_path}', adaptor '{adaptor_names}', " + f"backend '{backend}', prompt: '{prompts[0][:50]}...'" + ) + + if rouge_score < rouge_tol: + raise AssertionError( + f"ROUGE-L score {rouge_score} below tolerance {rouge_tol} " + f"for base '{base_path}', adaptor '{adaptor_names}', backend '{backend}', prompt: '{prompts[0][:50]}...'" + )