Add torchao quant (int4/int8/fp8) to llama models (#1341)

Co-authored-by: Lianmin Zheng <lianminzheng@gmail.com>

python/pyproject.toml

@@ -22,7 +22,7 @@ dependencies = [
 [project.optional-dependencies]
 srt = ["aiohttp", "decord", "fastapi", "hf_transfer", "huggingface_hub", "interegular",
        "packaging", "pillow", "psutil", "pydantic", "python-multipart",
-       "torch", "uvicorn", "uvloop", "zmq",
+       "torch", "torchao", "uvicorn", "uvloop", "zmq",
        "vllm==0.5.5", "outlines>=0.0.44"]
 openai = ["openai>=1.0", "tiktoken"]
 anthropic = ["anthropic>=0.20.0"]

python/sglang/srt/layers/torchao_utils.py

@@ -0,0 +1,36 @@
+"""
+Common utilities for torchao.
+"""
+
+import torch
+from torchao.quantization import (
+    int4_weight_only,
+    int8_dynamic_activation_int8_weight,
+    int8_weight_only,
+    quantize_,
+)
+
+
+def torchao_quantize_param_data(param, torchao_config):
+    dummy_linear = torch.nn.Linear(param.shape[1], param.shape[0], bias=False)
+    dummy_linear.weight = param
+    if "int8wo" in torchao_config:
+        quantize_(dummy_linear, int8_weight_only())
+    elif "int8dq" in torchao_config:
+        quantize_(dummy_linear, int8_dynamic_activation_int8_weight())
+    elif "int4wo" in torchao_config:
+        group_size = int(torchao_config.split("-")[-1])
+        assert group_size in [
+            32,
+            64,
+            128,
+            256,
+        ], f"int4wo groupsize needs to be one of [32, 64, 128, 256] but got {group_size}"
+        quantize_(dummy_linear, int4_weight_only(group_size=group_size))
+    elif "fp8wo" in torchao_config:
+        from torchao.quantization import float8_weight_only
+
+        # this requires newer hardware
+        # [rank0]: AssertionError: fp8e4nv data type is not supported on CUDA arch < 89
+        quantize_(dummy_linear, float8_weight_only())
+    return dummy_linear.weight

python/sglang/srt/model_executor/model_runner.py

@@ -97,6 +97,7 @@ class ModelRunner:
                 "disable_flashinfer_sampling": server_args.disable_flashinfer_sampling,
                 "triton_attention_reduce_in_fp32": server_args.triton_attention_reduce_in_fp32,
                 "enable_mla": server_args.enable_mla,
+                "torchao_config": server_args.torchao_config,
             }
         )
 

python/sglang/srt/models/llama.py

@@ -42,6 +42,8 @@ from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.sampler import Sampler
+from sglang.srt.layers.torchao_utils import torchao_quantize_param_data
+from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
@@ -299,6 +301,7 @@ class LlamaForCausalLM(nn.Module):
         super().__init__()
         self.config = config
         self.quant_config = quant_config
+        self.torchao_config = global_server_args_dict["torchao_config"]
         self.model = LlamaModel(config, quant_config=quant_config)
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
@@ -361,6 +364,25 @@ class LlamaForCausalLM(nn.Module):
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
 
+                if self.torchao_config:
+                    if name.endswith("proj.weight") and param.ndim == 2:
+                        params_dict[name] = torchao_quantize_param_data(
+                            param, self.torchao_config
+                        )
+
+        if self.torchao_config:
+            # quantizing the loaded, stacked params, e.g. "...qkv_proj"
+            stacked_params = set(entry[0] for entry in stacked_params_mapping)
+            for param_suffix in stacked_params:
+                for name in params_dict:
+                    if param_suffix in name:
+                        param = params_dict[name]
+                        params_dict[name] = torchao_quantize_param_data(
+                            param, self.torchao_config
+                        )
+
+            self.load_state_dict(params_dict, assign=True)
+
 
 class Phi3ForCausalLM(LlamaForCausalLM):
     pass

python/sglang/srt/server_args.py

@@ -95,6 +95,7 @@ class ServerArgs:
     disable_custom_all_reduce: bool = False
     enable_mixed_chunk: bool = False
     enable_torch_compile: bool = False
+    torchao_config: str = ""
     enable_p2p_check: bool = False
     enable_mla: bool = False
     triton_attention_reduce_in_fp32: bool = False
@@ -443,7 +444,13 @@ class ServerArgs:
         parser.add_argument(
             "--enable-torch-compile",
             action="store_true",
-            help="Optimize the model with torch.compile, experimental feature.",
+            help="Optimize the model with torch.compile. Experimental feature.",
+        )
+        parser.add_argument(
+            "--torchao-config",
+            type=str,
+            default=ServerArgs.torchao_config,
+            help="Optimize the model with torchao. Experimental feature. Current choices are: int8dq, int8wo, int4wo-<group_size>, fp8wo",
         )
         parser.add_argument(
             "--enable-p2p-check",

test/srt/test_eval_accuracy_mini.py

@@ -29,12 +29,12 @@ class TestEvalAccuracyMini(unittest.TestCase):
             base_url=self.base_url,
             model=self.model,
             eval_name="mmlu",
-            num_examples=32,
+            num_examples=64,
             num_threads=32,
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.6
+        assert metrics["score"] >= 0.65
 
 
 if __name__ == "__main__":

test/srt/test_moe_eval_accuracy_large.py

@@ -42,7 +42,7 @@ class TestEvalAccuracyLarge(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.62, f"{metrics}"
+        assert metrics["score"] >= 0.625, f"{metrics}"
 
     def test_human_eval(self):
         args = SimpleNamespace(
@@ -54,7 +54,7 @@ class TestEvalAccuracyLarge(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.42, f"{metrics}"
+        assert metrics["score"] >= 0.425, f"{metrics}"
 
     def test_mgsm_en(self):
         args = SimpleNamespace(
@@ -66,7 +66,7 @@ class TestEvalAccuracyLarge(unittest.TestCase):
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.62, f"{metrics}"
+        assert metrics["score"] >= 0.625, f"{metrics}"
 
 
 if __name__ == "__main__":

test/srt/test_torch_compile.py

@@ -22,7 +22,7 @@ class TestTorchCompile(unittest.TestCase):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=["--enable-torch-compile", "--disable-radix-cache"],
+            other_args=["--enable-torch-compile"],
         )
 
     @classmethod
@@ -34,12 +34,12 @@ class TestTorchCompile(unittest.TestCase):
             base_url=self.base_url,
             model=self.model,
             eval_name="mmlu",
-            num_examples=32,
+            num_examples=64,
             num_threads=32,
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.6
+        assert metrics["score"] >= 0.65
 
     def run_decode(self, max_new_tokens):
         response = requests.post(

test/srt/test_torchao.py

@@ -0,0 +1,73 @@
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+from sglang.srt.utils import kill_child_process
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
+
+
+class TestTorchCompile(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_MODEL_NAME_FOR_TEST
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=["--torchao-config", "int4wo-128"],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_child_process(cls.process.pid)
+
+    def test_mmlu(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mmlu",
+            num_examples=64,
+            num_threads=32,
+        )
+
+        metrics = run_eval(args)
+        assert metrics["score"] >= 0.65
+
+    def run_decode(self, max_new_tokens):
+        response = requests.post(
+            self.base_url + "/generate",
+            json={
+                "text": "The capital of France is",
+                "sampling_params": {
+                    "temperature": 0,
+                    "max_new_tokens": max_new_tokens,
+                },
+                "ignore_eos": True,
+            },
+        )
+        return response.json()
+
+    def test_throughput(self):
+        import time
+
+        max_tokens = 256
+
+        tic = time.time()
+        res = self.run_decode(max_tokens)
+        tok = time.time()
+        print(res["text"])
+        throughput = max_tokens / (tok - tic)
+        print(f"Throughput: {throughput} tokens/s")
+        assert throughput >= 210
+
+
+if __name__ == "__main__":
+    unittest.main()

test/srt/test_triton_attn_backend.py

@@ -32,12 +32,12 @@ class TestTritonAttnBackend(unittest.TestCase):
             base_url=self.base_url,
             model=self.model,
             eval_name="mmlu",
-            num_examples=32,
+            num_examples=64,
             num_threads=32,
         )
 
         metrics = run_eval(args)
-        assert metrics["score"] >= 0.6
+        assert metrics["score"] >= 0.65
 
 
 if __name__ == "__main__":