Fix bugs (fp8 checkpoints, triton cache manager)

README.md

@@ -70,11 +70,6 @@ docker run --gpus all \
 ```
 
 ### Common Notes
-- If you see errors from the Triton compiler, please install the [Triton Nightly](https://triton-lang.org/main/getting-started/installation.html) by
-```
-pip uninstall -y triton triton-nightly
-pip install -U --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/ triton-nightly
-```
 - If you cannot install FlashInfer, check out its [installation](https://docs.flashinfer.ai/installation.html#) page. If you still cannot install it, you can use the slower Triton kernels by adding `--disable-flashinfer` when launching the server.
 - If you only need to use the OpenAI backend, you can avoid installing other dependencies by using `pip install "sglang[openai]"`.
 
@@ -157,6 +152,7 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 ```
 - If the model does not have a template in the Hugging Face tokenizer, you can specify a [custom chat template](docs/custom_chat_template.md).
 - To enable fp8 quantization, you can add `--quantization fp8` on a fp16 checkpoint or directly load a fp8 checkpoint without specifying any arguments.
+- To enable experimental torch.compile support, you can add `--enable-torch-compile`. It accelerates small models on small batch sizes.
 
 ### Supported Models
 

python/sglang/srt/managers/controller/model_runner.py

@@ -30,9 +30,11 @@
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import (
     get_available_gpu_memory,
+    is_llama3_405b_fp8,
     is_multimodal_model,
     monkey_patch_vllm_dummy_weight_loader,
     monkey_patch_vllm_p2p_access_check,
+    monkey_patch_vllm_qvk_linear_loader,
 )
 
 logger = logging.getLogger("srt.model_runner")
@@ -118,6 +120,13 @@ def load_model(self):
             seed=42,
             skip_tokenizer_init=True,
         )
+
+        if is_llama3_405b_fp8(self.model_config):
+            # A temporary hack to fix the num_heads for meta-llama/Meta-Llama-3.1-405B-FP8 checkpoints
+            self.model_config.hf_config.num_key_value_heads = 8
+            vllm_model_config.hf_config.num_key_value_heads = 8
+            monkey_patch_vllm_qvk_linear_loader()
+
         self.dtype = vllm_model_config.dtype
         if self.model_config.model_overide_args is not None:
             vllm_model_config.hf_config.update(self.model_config.model_overide_args)

python/sglang/srt/server.py

@@ -202,15 +202,12 @@ def launch_server(
             "reinstall the latest version by following the instructions "
             "at https://docs.flashinfer.ai/installation.html.",
         )
-
-    if server_args.tp_size // server_args.dp_size > 1:
+    if server_args.tp_size * server_args.dp_size > 1:
         # FIXME: remove this after https://github.com/triton-lang/triton/pull/4295 is used as a dependency.
         maybe_set_triton_cache_manager()
-
     if server_args.chat_template:
         # TODO: replace this with huggingface transformers template
         load_chat_template_for_openai_api(server_args.chat_template)
-
     if server_args.enable_torch_compile:
         _set_torch_compile_config()
 

python/sglang/srt/utils.py

@@ -21,6 +21,7 @@
 from fastapi.responses import JSONResponse
 from packaging import version as pkg_version
 from starlette.middleware.base import BaseHTTPMiddleware
+from torch.nn.parameter import Parameter
 from triton.runtime.cache import (
     FileCacheManager,
     default_cache_dir,
@@ -471,7 +472,7 @@ def maybe_set_triton_cache_manager() -> None:
     cache_manger = os.environ.get("TRITON_CACHE_MANAGER", None)
     if cache_manger is None:
         manager = "sglang.srt.utils:CustomCacheManager"
-        logger.info("Setting Triton cache manager to: %s", manager)
+        logger.debug("Setting Triton cache manager to: %s", manager)
         os.environ["TRITON_CACHE_MANAGER"] = manager
 
 
@@ -615,3 +616,51 @@ def set_ulimit(target_soft_limit=65535):
             resource.setrlimit(resource_type, (target_soft_limit, current_hard))
         except ValueError as e:
             logger.warn(f"Fail to set RLIMIT_NOFILE: {e}")
+
+
+def is_llama3_405b_fp8(model_config):
+    """Return whether the model is meta-llama/Meta-Llama-3.1-405B-FP8 with 16 kv heads."""
+    if (
+        model_config.hf_config.architectures[0] == "LlamaForCausalLM"
+        and model_config.hf_config.hidden_size == 16384
+        and model_config.hf_config.intermediate_size == 53248
+        and model_config.hf_config.num_hidden_layers == 126
+        and model_config.hf_config.num_key_value_heads == 16
+        and model_config.hf_config.quantization_config["quant_method"] == "fbgemm_fp8"
+    ):
+        return True
+    return False
+
+
+def monkey_patch_vllm_qvk_linear_loader():
+    """A temporary hack to fix the num_heads for meta-llama/Meta-Llama-3.1-405B-FP8 checkpoints."""
+    from vllm.model_executor.layers.linear import QKVParallelLinear
+
+    origin_weight_loader = QKVParallelLinear.weight_loader
+
+    def get_original_weight(loaded_weight, head_dim):
+        n_kv_head = loaded_weight.shape[0] // (2 * head_dim)
+        dim = loaded_weight.shape[1]
+        for i in range(n_kv_head):
+            loaded_weight[i * head_dim : (i + 1) * head_dim, :] = loaded_weight[
+                2 * i * head_dim : (2 * i + 1) * head_dim, :
+            ]
+        original_kv_weight = loaded_weight[: n_kv_head * head_dim, :]
+        assert original_kv_weight.shape == (n_kv_head * head_dim, dim)
+        return original_kv_weight
+
+    def weight_loader_srt(
+        self,
+        param: Parameter,
+        loaded_weight: torch.Tensor,
+        loaded_shard_id: Optional[str] = None,
+    ):
+        if (
+            loaded_shard_id in ["k", "v"]
+            and loaded_weight.shape[0] == self.head_size * self.total_num_kv_heads * 2
+        ):
+            loaded_weight = get_original_weight(loaded_weight, self.head_size)
+
+        origin_weight_loader(self, param, loaded_weight, loaded_shard_id)
+
+    setattr(QKVParallelLinear, "weight_loader", weight_loader_srt)