[Feature][kernel] tensor parallelism with bitsandbytes quantization (#8434)

tests/quantization/test_bitsandbytes.py

@@ -64,6 +64,24 @@ def test_load_8bit_bnb_model(hf_runner, vllm_runner, example_prompts,
                              model_name)
 
 
+@pytest.mark.skipif(torch.cuda.device_count() < 2,
+                    reason='Test requires at least 2 GPUs.')
+@pytest.mark.skipif(not is_quant_method_supported("bitsandbytes"),
+                    reason='bitsandbytes is not supported on this GPU type.')
+@pytest.mark.parametrize("model_name, description", models_4bit_to_test)
+@fork_new_process_for_each_test
+def test_load_tp_4bit_bnb_model(hf_runner, vllm_runner, example_prompts,
+                                model_name, description) -> None:
+
+    hf_model_kwargs = {"load_in_4bit": True}
+    validate_generated_texts(hf_runner,
+                             vllm_runner,
+                             example_prompts[:1],
+                             model_name,
+                             hf_model_kwargs,
+                             vllm_tp_size=2)
+
+
 def log_generated_texts(prompts, outputs, runner_name):
     logged_texts = []
     for i, (_, generated_text) in enumerate(outputs):
@@ -80,22 +98,21 @@ def validate_generated_texts(hf_runner,
                              vllm_runner,
                              prompts,
                              model_name,
-                             hf_model_kwargs=None):
+                             hf_model_kwargs=None,
+                             vllm_tp_size=1):
 
     # NOTE: run vLLM first, as it requires a clean process
     # when using distributed inference
-
-    #Run with vLLM runner
     with vllm_runner(model_name,
                      quantization='bitsandbytes',
                      load_format='bitsandbytes',
+                     tensor_parallel_size=vllm_tp_size,
                      enforce_eager=True,
                      gpu_memory_utilization=0.8) as llm:
         vllm_outputs = llm.generate_greedy(prompts, 8)
         vllm_logs = log_generated_texts(prompts, vllm_outputs, "VllmRunner")
 
     # Clean up the GPU memory for the next test
-    torch.cuda.synchronize()
     gc.collect()
     torch.cuda.empty_cache()
 
@@ -108,7 +125,6 @@ def validate_generated_texts(hf_runner,
         hf_logs = log_generated_texts(prompts, hf_outputs, "HfRunner")
 
     # Clean up the GPU memory for the next test
-    torch.cuda.synchronize()
     gc.collect()
     torch.cuda.empty_cache()
 

vllm/config.py

@@ -393,12 +393,6 @@ class ModelConfig:
                 "Pipeline parallelism is only supported for the following "
                 f" architectures: {_PP_SUPPORTED_MODELS}.")
 
-        if self.quantization == "bitsandbytes" and (
-                parallel_config.tensor_parallel_size > 1
-                or parallel_config.pipeline_parallel_size > 1):
-            raise ValueError(
-                "BitAndBytes quantization with TP or PP is not supported yet.")
-
         # Remove the constraint after the bitsandbytes issue is fixed:
         # https://github.com/bitsandbytes-foundation/bitsandbytes/issues/1308
         if self.quantization == "bitsandbytes" and self.enforce_eager is False:

vllm/model_executor/layers/linear.py

@@ -530,8 +530,11 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
             param_data = param_data.narrow(output_dim, shard_offset,
                                            shard_size)
             start_idx = tp_rank * shard_size
-            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
-                                                 shard_size)
+            # bitsandbytes loads the weights of the specific portion
+            # no need to narrow here
+            if not use_bitsandbytes_4bit:
+                loaded_weight = loaded_weight.narrow(output_dim, start_idx,
+                                                     shard_size)
         # Special case for AQLM codebooks.
         elif is_metadata:
             # metadata indicates fixed size concatenated along dim 0
@@ -899,8 +902,13 @@ class QKVParallelLinear(ColumnParallelLinear):
             else:
                 shard_id = tp_rank // self.num_kv_head_replicas
             start_idx = shard_id * shard_size
-            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
-                                                 shard_size)
+
+            # bitsandbytes loads the weights of the specific portion
+            # no need to narrow here
+            if not use_bitsandbytes_4bit:
+                loaded_weight = loaded_weight.narrow(output_dim, start_idx,
+                                                     shard_size)
+
         # Special case for for AQLM codebooks.
         elif is_metadata:
             # metadata indicates fixed size concatenated along dim 0
@@ -1000,6 +1008,7 @@ class RowParallelLinear(LinearBase):
         tp_rank = get_tensor_model_parallel_rank()
         tp_size = get_tensor_model_parallel_world_size()
         input_dim = getattr(param, "input_dim", None)
+        use_bitsandbytes_4bit = getattr(param, "use_bitsandbytes_4bit", False)
 
         # Special case for GGUF
         is_gguf_weight = getattr(param, "is_gguf_weight", False)
@@ -1015,7 +1024,9 @@ class RowParallelLinear(LinearBase):
             param.materialize(tuple(weight_shape), dtype=loaded_weight.dtype)
 
         param_data = param.data
-        if input_dim is not None:
+        # bitsandbytes loads the weights of the specific portion
+        # no need to narrow here
+        if input_dim is not None and not use_bitsandbytes_4bit:
             shard_size = param_data.shape[input_dim]
             start_idx = tp_rank * shard_size
             loaded_weight = loaded_weight.narrow(input_dim, start_idx,

vllm/model_executor/model_loader/loader.py

@@ -22,6 +22,8 @@ from transformers.utils import SAFE_WEIGHTS_INDEX_NAME
 from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoadFormat,
                          LoRAConfig, ModelConfig, MultiModalConfig,
                          ParallelConfig, SchedulerConfig)
+from vllm.distributed import (get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size)
 from vllm.envs import VLLM_USE_MODELSCOPE
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization.base_config import (
@@ -689,6 +691,8 @@ class ShardedStateLoader(BaseModelLoader):
 class BitsAndBytesModelLoader(BaseModelLoader):
     """Model loader to load model weights with BitAndBytes quantization."""
 
+    # TODO: these module names are for Llama only,
+    # change so that it works with other models as well
     default_target_modules = [
         "gate_proj", "down_proj", "up_proj", "q_proj", "k_proj", "v_proj",
         "o_proj"
@@ -911,13 +915,44 @@ class BitsAndBytesModelLoader(BaseModelLoader):
     def _unquantized_generator(self, hf_weights_files, use_safetensors,
                                quant_state_dict) -> Generator:
         from bitsandbytes.functional import quantize_4bit
+        tp_size = get_tensor_model_parallel_world_size()
+        tp_rank = get_tensor_model_parallel_rank()
+
         for weight_name, weight_tensor in self._hf_weight_iter(
                 hf_weights_files, use_safetensors):
             if any(target_module in weight_name
                    for target_module in self.target_modules):
                 weight_name = weight_name.replace(".weight", ".qweight")
+
+                # weight partitions of different modules occur at
+                # different dimensions
+                # TODO: these module names are for Llama only,
+                # change so that it works with other models as well
+                if 'down_proj' in weight_name or 'o_proj' in weight_name:
+                    total_size = weight_tensor.size(-1)
+                    start_index = total_size // tp_size * tp_rank
+                    end_index = total_size // tp_size * (tp_rank + 1)
+                    weight_sub_tensor = weight_tensor[...,
+                                                      start_index:end_index]
+
+                else:
+                    total_size = weight_tensor.size(0)
+                    start_index = total_size // tp_size * tp_rank
+                    end_index = total_size // tp_size * (tp_rank + 1)
+                    weight_sub_tensor = weight_tensor[start_index:end_index,
+                                                      ...]
+
                 # bitsandbytes requires data in GPU
-                loaded_weight = weight_tensor.cuda().data
+                if weight_sub_tensor.is_cuda:
+                    loaded_weight = weight_sub_tensor
+                else:
+                    loaded_weight = weight_sub_tensor.cuda()
+
+                # remove the following after the issue is fixed:
+                # https://github.com/bitsandbytes-foundation/bitsandbytes/issues/1342
+                if loaded_weight.is_contiguous() is False:
+                    loaded_weight = loaded_weight.contiguous()
+
                 with set_default_torch_dtype(torch.float32):
                     processed_weight, quant_state = quantize_4bit(
                         loaded_weight,
@@ -958,6 +993,13 @@ class BitsAndBytesModelLoader(BaseModelLoader):
                     f"BitsAndBytes loader does not support {quant_method} "
                     "quantization")
 
+        # The quant_states in pre_quantized models cannot work with a split
+        # weight tensor. So TP does not work with pre_quantized bnb models.
+        if pre_quant and get_tensor_model_parallel_world_size() > 1:
+            raise ValueError(
+                "Prequant BitsAndBytes models with TP is not supported."
+                "Please try with PP.")
+
         load_8bit = False
         if pre_quant:
             load_8bit = quant_config.get('load_in_8bit', False)