[inference] Refactor inference architecture (#5057)

* [inference] support only TP (#4998) * support only tp * enable tp * add support for bloom (#5008) * [refactor] refactor gptq and smoothquant llama (#5012) * refactor gptq and smoothquant llama * fix import error * fix linear import torch-int * fix smoothquant llama import error * fix import accelerate error * fix bug * fix import smooth cuda * fix smoothcuda * [Inference Refactor] Merge chatglm2 with pp and tp (#5023) merge chatglm with pp and tp * [Refactor] remove useless inference code (#5022) * remove useless code * fix quant model * fix test import bug * mv original inference legacy * fix chatglm2 * [Refactor] refactor policy search and quant type controlling in inference (#5035) * [Refactor] refactor policy search and quant type controling in inference * [inference] update readme (#5051) * update readme * update readme * fix architecture * fix table * fix table * [inference] udpate example (#5053) * udpate example * fix run.sh * fix rebase bug * fix some errors * update readme * add some features * update interface * update readme * update benchmark * add requirements-infer --------- Co-authored-by: Bin Jia <45593998+FoolPlayer@users.noreply.github.com> Co-authored-by: Zhongkai Zhao <kanezz620@gmail.com>
2025-10-01 15:18:51 +00:00 · 2023-11-19 21:05:05 +08:00
parent bc09b95f50
commit fd6482ad8c
115 changed files with 6027 additions and 1431 deletions
--- a/examples/inference/hybrid_smoothquant_llama.py
+++ b/examples/inference/hybrid_smoothquant_llama.py
@@ -0,0 +1,69 @@
+import argparse
+
+import torch
+import torch.distributed as dist
+
+import colossalai
+from colossalai.inference import CaiInferEngine
+from colossalai.inference.quant.smoothquant.models.llama import SmoothLlamaForCausalLM
+from colossalai.logging import disable_existing_loggers
+from colossalai.testing import spawn
+
+
+@torch.no_grad()
+def run_llama_inference(args):
+    quantized_model_dir = args.quantized_path
+    max_batch_size = args.max_batch_size
+    max_input_len = args.max_input_len
+    max_output_len = args.max_output_len
+    micro_batch_size = args.micro_batch_size
+
+    def data_gen():
+        input_ids = torch.tensor([[15496, 11, 616, 3290, 318, 13779, 318, 13779]], dtype=torch.int64)
+        attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1]], dtype=torch.int64)
+        return dict(input_ids=input_ids, attention_mask=attention_mask)
+
+    inputs = data_gen()
+    for k, v in inputs.items():
+        if torch.is_tensor(v) or "Tensor" in v.__class__.__name__:
+            new_shape = [1] * v.dim()
+            new_shape[0] = 16
+            inputs[k] = v.to("cuda").repeat(*new_shape)
+
+    model = SmoothLlamaForCausalLM.from_quantized(quantized_model_dir, model_basename="llama-7b")
+    model = model.cuda()
+
+    engine = CaiInferEngine(
+        tp_size=2,
+        pp_size=2,
+        model=model,
+        max_batch_size=max_batch_size,
+        max_input_len=max_input_len,
+        max_output_len=max_output_len,
+        micro_batch_size=micro_batch_size,
+        quant="smoothquant",
+    )
+
+    output = engine.generate(inputs)
+    if dist.get_rank() == 0:
+        assert len(output[0]) == 32, f"{len(output)}, {32}"
+
+
+def run_smoothquant_inference(rank, world_size, port, args):
+    disable_existing_loggers()
+    colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
+    run_llama_inference(args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-q", "--quantized_path", type=str, help="Model path", required=True)
+    parser.add_argument("--tp_size", type=int, default=2, help="Tensor parallel size")
+    parser.add_argument("--pp_size", type=int, default=2, help="Pipeline parallel size")
+    parser.add_argument("--max_batch_size", type=int, default=4, help="Maximum batch size")
+    parser.add_argument("--micro_batch_size", type=int, default=4, help="Micro batch size")
+    parser.add_argument("--max_input_len", type=int, default=32, help="Maximum input length")
+    parser.add_argument("--max_output_len", type=int, default=32, help="Maximum output length")
+
+    args = parser.parse_args()
+    spawn(run_smoothquant_inference, args.tp_size * args.pp_size, args=args)