Enable more models to inference based on LoRA (#3382)

Co-authored-by: Antoni Baum <[email protected]>

csrc/punica/bgmv/bgmv_config.h

@@ -16,21 +16,26 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 512) \
     f(in_T, out_T, W_T, narrow, 768) \
     f(in_T, out_T, W_T, narrow, 1024) \
+    f(in_T, out_T, W_T, narrow, 1152) \
     f(in_T, out_T, W_T, narrow, 1280) \
+    f(in_T, out_T, W_T, narrow, 1536) \
     f(in_T, out_T, W_T, narrow, 1728) \
     f(in_T, out_T, W_T, narrow, 1792) \
     f(in_T, out_T, W_T, narrow, 2048) \
+    f(in_T, out_T, W_T, narrow, 2304) \
     f(in_T, out_T, W_T, narrow, 2560) \
     f(in_T, out_T, W_T, narrow, 2752) \
     f(in_T, out_T, W_T, narrow, 2816) \
     f(in_T, out_T, W_T, narrow, 3072) \
     f(in_T, out_T, W_T, narrow, 3456) \
     f(in_T, out_T, W_T, narrow, 3584) \
     f(in_T, out_T, W_T, narrow, 4096) \
+    f(in_T, out_T, W_T, narrow, 4608) \
     f(in_T, out_T, W_T, narrow, 5120) \
     f(in_T, out_T, W_T, narrow, 5504) \
     f(in_T, out_T, W_T, narrow, 5632) \
     f(in_T, out_T, W_T, narrow, 6144) \
+    f(in_T, out_T, W_T, narrow, 6848) \
     f(in_T, out_T, W_T, narrow, 6912) \
     f(in_T, out_T, W_T, narrow, 7168) \
     f(in_T, out_T, W_T, narrow, 8192) \
@@ -45,6 +50,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 20480) \
     f(in_T, out_T, W_T, narrow, 22016) \
     f(in_T, out_T, W_T, narrow, 24576) \
+    f(in_T, out_T, W_T, narrow, 27392) \
     f(in_T, out_T, W_T, narrow, 28672) \
     f(in_T, out_T, W_T, narrow, 32000) \
     f(in_T, out_T, W_T, narrow, 32256) \

tests/lora/conftest.py

@@ -134,6 +134,16 @@ def gemma_lora_files():
     return snapshot_download(repo_id="wskwon/gemma-7b-test-lora")
 
 
+@pytest.fixture(scope="session")
+def chatglm3_lora_files():
+    return snapshot_download(repo_id="jeeejeee/chatglm3-text2sql-spider")
+
+
+@pytest.fixture(scope="session")
+def baichuan_lora_files():
+    return snapshot_download(repo_id="jeeejeee/baichuan7b-text2sql-spider")
+
+
 @pytest.fixture
 def llama_2_7b_engine_extra_embeddings() -> nn.Module:
     cleanup()

tests/lora/test_baichuan.py

@@ -0,0 +1,108 @@
+import pytest
+
+import vllm
+from vllm.lora.request import LoRARequest
+
+from .conftest import cleanup
+
+MODEL_PATH = "baichuan-inc/Baichuan-7B"
+
+PROMPT_TEMPLATE = """I want you to act as a SQL terminal in front of an example database, you need only to return the sql command to me.Below is an instruction that describes a task, Write a response that appropriately completes the request.\n"\n##Instruction:\nconcert_singer contains tables such as stadium, singer, concert, singer_in_concert. Table stadium has columns such as Stadium_ID, Location, Name, Capacity, Highest, Lowest, Average. Stadium_ID is the primary key.\nTable singer has columns such as Singer_ID, Name, Country, Song_Name, Song_release_year, Age, Is_male. Singer_ID is the primary key.\nTable concert has columns such as concert_ID, concert_Name, Theme, Stadium_ID, Year. concert_ID is the primary key.\nTable singer_in_concert has columns such as concert_ID, Singer_ID. concert_ID is the primary key.\nThe Stadium_ID of concert is the foreign key of Stadium_ID of stadium.\nThe Singer_ID of singer_in_concert is the foreign key of Singer_ID of singer.\nThe concert_ID of singer_in_concert is the foreign key of concert_ID of concert.\n\n###Input:\n{query}\n\n###Response:"""  # noqa: E501
+
+
+def do_sample(llm, lora_path: str, lora_id: int) -> str:
+    prompts = [
+        PROMPT_TEMPLATE.format(query="How many singers do we have?"),
+        PROMPT_TEMPLATE.format(
+            query=
+            "What is the average, minimum, and maximum age of all singers from France?"  # noqa: E501
+        ),
+        PROMPT_TEMPLATE.format(
+            query=
+            "Show name, country, age for all singers ordered by age from the oldest to the youngest."  # noqa: E501
+        ),
+    ]
+    print(prompts)
+    sampling_params = vllm.SamplingParams(temperature=0, max_tokens=256)
+    outputs = llm.generate(
+        prompts,
+        sampling_params,
+        lora_request=LoRARequest(str(lora_id), lora_id, lora_path)
+        if lora_id else None)
+    # Print the outputs.
+    generated_texts = []
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text.strip()
+        generated_texts.append(generated_text)
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+    return generated_texts
+
+
+def test_baichuan_lora(baichuan_lora_files):
+    llm = vllm.LLM(MODEL_PATH,
+                   max_model_len=1024,
+                   enable_lora=True,
+                   max_loras=4,
+                   max_lora_rank=64,
+                   trust_remote_code=True)
+
+    expected_lora_output = [
+        "SELECT count(*) FROM singer",
+        "SELECT avg(age) ,  min(age) ,  max(age) FROM singer WHERE Country  =  'France'",  # noqa: E501
+        "SELECT name ,  country ,  age FROM singer ORDER BY age ASC",
+    ]
+
+    output1 = do_sample(llm, baichuan_lora_files, lora_id=1)
+    for i in range(len(expected_lora_output)):
+        assert output1[i] == expected_lora_output[i]
+    output2 = do_sample(llm, baichuan_lora_files, lora_id=2)
+    for i in range(len(expected_lora_output)):
+        assert output2[i] == expected_lora_output[i]
+
+
+@pytest.mark.skip("Requires multiple GPUs")
+def test_llama_tensor_parallel_equality(baichuan_lora_files):
+    # Cannot use as it will initialize torch.cuda too early...
+    # if torch.cuda.device_count() < 4:
+    #     pytest.skip(f"Not enough GPUs for tensor parallelism {4}")
+
+    llm_tp1 = vllm.LLM(MODEL_PATH,
+                       enable_lora=True,
+                       max_num_seqs=16,
+                       max_loras=4,
+                       max_lora_rank=64,
+                       tensor_parallel_size=1,
+                       trust_remote_code=True)
+    output_tp1 = do_sample(llm_tp1, baichuan_lora_files, lora_id=1)
+
+    del llm_tp1
+    cleanup()
+
+    llm_tp2 = vllm.LLM(MODEL_PATH,
+                       enable_lora=True,
+                       max_num_seqs=16,
+                       max_loras=4,
+                       max_lora_rank=64,
+                       tensor_parallel_size=2,
+                       trust_remote_code=True)
+    output_tp2 = do_sample(llm_tp2, baichuan_lora_files, lora_id=2)
+
+    del llm_tp2
+    cleanup()
+
+    assert output_tp1 == output_tp2
+
+    llm_tp4 = vllm.LLM(MODEL_PATH,
+                       enable_lora=True,
+                       max_num_seqs=16,
+                       max_loras=4,
+                       max_lora_rank=64,
+                       tensor_parallel_size=4,
+                       trust_remote_code=True)
+    output_tp4 = do_sample(llm_tp4, baichuan_lora_files, lora_id=2)
+
+    del llm_tp4
+    cleanup()
+
+    assert output_tp1 == output_tp4

tests/lora/test_chatglm3.py

@@ -0,0 +1,57 @@
+import vllm
+from vllm.lora.request import LoRARequest
+
+MODEL_PATH = "THUDM/chatglm3-6b"
+
+PROMPT_TEMPLATE = """I want you to act as a SQL terminal in front of an example database, you need only to return the sql command to me.Below is an instruction that describes a task, Write a response that appropriately completes the request.\n"\n##Instruction:\nconcert_singer contains tables such as stadium, singer, concert, singer_in_concert. Table stadium has columns such as Stadium_ID, Location, Name, Capacity, Highest, Lowest, Average. Stadium_ID is the primary key.\nTable singer has columns such as Singer_ID, Name, Country, Song_Name, Song_release_year, Age, Is_male. Singer_ID is the primary key.\nTable concert has columns such as concert_ID, concert_Name, Theme, Stadium_ID, Year. concert_ID is the primary key.\nTable singer_in_concert has columns such as concert_ID, Singer_ID. concert_ID is the primary key.\nThe Stadium_ID of concert is the foreign key of Stadium_ID of stadium.\nThe Singer_ID of singer_in_concert is the foreign key of Singer_ID of singer.\nThe concert_ID of singer_in_concert is the foreign key of concert_ID of concert.\n\n###Input:\n{query}\n\n###Response:"""  # noqa: E501
+
+
+def do_sample(llm, lora_path: str, lora_id: int) -> str:
+    prompts = [
+        PROMPT_TEMPLATE.format(query="How many singers do we have?"),
+        PROMPT_TEMPLATE.format(
+            query=
+            "What is the average, minimum, and maximum age of all singers from France?"  # noqa: E501
+        ),
+        PROMPT_TEMPLATE.format(
+            query=
+            "Show name, country, age for all singers ordered by age from the oldest to the youngest."  # noqa: E501
+        ),
+    ]
+    print(prompts)
+    sampling_params = vllm.SamplingParams(temperature=0, max_tokens=32)
+    outputs = llm.generate(
+        prompts,
+        sampling_params,
+        lora_request=LoRARequest(str(lora_id), lora_id, lora_path)
+        if lora_id else None)
+    # Print the outputs.
+    generated_texts = []
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text.strip()
+        generated_texts.append(generated_text)
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+    return generated_texts
+
+
+def test_chatglm3_lora(chatglm3_lora_files):
+    llm = vllm.LLM(MODEL_PATH,
+                   max_model_len=1024,
+                   enable_lora=True,
+                   max_loras=4,
+                   max_lora_rank=64,
+                   trust_remote_code=True)
+
+    expected_lora_output = [
+        "SELECT count(*) FROM singer",
+        "SELECT avg(age) ,  min(age) ,  max(age) FROM singer WHERE country  =  'France'",  # noqa: E501
+        "SELECT name ,  country ,  age FROM singer ORDER BY age",
+    ]
+
+    output1 = do_sample(llm, chatglm3_lora_files, lora_id=1)
+    for i in range(len(expected_lora_output)):
+        assert output1[i] == expected_lora_output[i]
+    output2 = do_sample(llm, chatglm3_lora_files, lora_id=2)
+    for i in range(len(expected_lora_output)):
+        assert output2[i] == expected_lora_output[i]

tests/lora/test_layers.py

@@ -8,12 +8,16 @@
 import torch.nn.functional as F
 
 from vllm.config import LoRAConfig
+# yapf conflicts with isort for this block
+# yapf: disable
 from vllm.lora.layers import (BaseLayerWithLoRA, ColumnParallelLinearWithLoRA,
                               LogitsProcessorWithLoRA, LoRAMapping,
                               MergedColumnParallelLinearWithLoRA,
+                              MergedQKVParallelLinearWithLora,
                               QKVParallelLinearWithLora,
                               RowParallelLinearWithLoRA,
                               VocabParallelEmbeddingWithLoRA)
+# yapf: enable
 from vllm.lora.models import (LoRALayerWeights, PackedLoRALayerWeights,
                               convert_mapping)
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
@@ -93,8 +97,7 @@ def populate_loras(
     lora_dict: Dict[int, LoRALayerWeights] = dict()
 
     # Dictionary that maps the lora ID to the
-    # corresponding subloras. Only useful when
-    # repeats > 1.
+    # corresponding subloras.
     sublora_dict: Dict[int, List[LoRALayerWeights]] = dict()
 
     for slot_idx, lora_id in enumerate(id_to_index):
@@ -607,7 +610,7 @@ def create_random_linear_parallel_layer():
 
 @torch.inference_mode()
 @pytest.mark.parametrize("num_loras", [1, 2, 4, 8])
-@pytest.mark.parametrize("repeats", [2, 3])
+@pytest.mark.parametrize("repeats", [1, 2, 3])
 @pytest.mark.parametrize("device", CUDA_DEVICES)
 def test_column_parallel_packed(dist_init, num_loras, repeats, device) -> None:
 
@@ -623,6 +626,10 @@ def create_column_parallel_packed_layer():
                                                 bias=False)
             linear.weight.data = torch.rand_like(linear.weight.data)
             lora_linear = MergedColumnParallelLinearWithLoRA(linear)
+        elif repeats == 3:
+            linear = QKVParallelLinear(4096, 64, 32, bias=False)
+            linear.weight.data = torch.rand_like(linear.weight.data)
+            lora_linear = MergedQKVParallelLinearWithLora(linear)
         else:
             linear = QKVParallelLinear(4096, 64, 32, bias=False)
             linear.weight.data = torch.rand_like(linear.weight.data)

tests/lora/test_punica.py

@@ -43,9 +43,10 @@ def _lora_ref_impl(
 
 
 H1 = H2 = [
-    128, 256, 512, 1024, 1280, 2048, 2560, 2752, 3072, 3456, 3584, 4096, 5120,
-    5504, 5632, 6144, 6912, 7168, 8192, 9216, 10240, 11008, 13824, 14336,
-    22016, 24576, 32000, 32256, 32512, 32768, 33024
+    128, 256, 512, 1024, 1152, 1280, 1536, 2048, 2304, 2560, 2752, 3072, 3456,
+    3584, 4096, 4608, 5120, 5504, 5632, 6144, 6848, 6912, 7168, 8192, 9216,
+    10240, 11008, 13824, 14336, 22016, 24576, 27392, 32000, 32256, 32512,
+    32768, 33024
 ]
 SEED = [0xabcdabcd987]