fix typos (#3684)

fastdeploy/cache_manager/cache_transfer_manager.py

@@ -57,7 +57,7 @@ def parse_args():
         "--protocol",
         type=str,
         default="ipc",
-        help="cache transfer protocol, only surport ipc now",
+        help="cache transfer protocol, only support ipc now",
     )
     parser.add_argument("--enable_splitwise", type=int, default=0, help="enable splitwise ")
     parser.add_argument("--cache_queue_port", type=int, default=9923, help="cache queue port")

fastdeploy/config.py

@@ -257,7 +257,7 @@ class ParallelConfig:
         self.sequence_parallel = False  # Whether to enable sequence parallelism.
         self.use_ep = False  # Whether to enable Expert Parallelism
         self.moe_phase = MoEPhase("prefill")  # Generation phase
-        self.msg_queue_id = 1  # mesage queue id
+        self.msg_queue_id = 1  # message queue id
 
         self.tensor_parallel_rank = 0  # TP rank ID
         self.tensor_parallel_size = 1  # TP degree
@@ -549,7 +549,7 @@ class GraphOptimizationConfig:
             It requires that all input buffers have fixed addresses, and all
             splitting ops write their outputs to input buffers.
             - With dyncmic graph backend: ...
-            - With static grpah backend: WIP
+            - With static graph backend: WIP
         """
         self.sot_warmup_sizes: list[int] = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 16, 32, 64, 128]
         """  Number of warmup runs for SOT warmup. """

fastdeploy/engine/args_utils.py

@@ -531,7 +531,7 @@ class EngineArgs:
             "--quantization",
             type=str,
             default=EngineArgs.quantization,
-            help="Quantization name for the model, currentlly support "
+            help="Quantization name for the model, currently support "
             "'wint8', 'wint4',"
             "default is None. The priority of this configuration "
             "is lower than that of the config file. "
@@ -829,7 +829,7 @@ class EngineArgs:
         scheduler_group.add_argument(
             "--scheduler-topic",
             default=EngineArgs.scheduler_topic,
-            help=f"Topic of scheduler. Defaule is {EngineArgs.scheduler_topic}. (global)",
+            help=f"Topic of scheduler. Default is {EngineArgs.scheduler_topic}. (global)",
         )
         scheduler_group.add_argument(
             "--scheduler-min-load-score",

fastdeploy/engine/common_engine.py

@@ -644,13 +644,13 @@ class EngineSevice:
                     self.zmq_server.send_multipart(request_id, [error_result])
             except Exception as e:
                 llm_logger.error(
-                    f"Error happend while receving new request from zmq, details={e}, "
+                    f"Error happend while receiving new request from zmq, details={e}, "
                     f"traceback={traceback.format_exc()}"
                 )
 
     def _zmq_send_generated_tokens(self):
         """
-        Recieve output for zmq
+        Receive output for zmq
         """
         while self.running:
             try:

fastdeploy/engine/sched/resource_manager_v1.py

@@ -458,7 +458,7 @@ class ResourceManagerV1(ResourceManager):
 
     def _free_blocks(self, request: Request):
         if self.config.cache_config.enable_prefix_caching:
-            # TODO(chengyanfu): support cache ouput blocks for prefix caching
+            # TODO(chengyanfu): support cache output blocks for prefix caching
             if request.get("prefill_block_num", None) is None:
                 leaf_node = self.cache_manager.req_leaf_map[request.request_id]
                 self.cache_manager.decrease_request_share_count(request.request_id)

fastdeploy/entrypoints/llm.py

@@ -112,7 +112,7 @@ class LLM:
 
     def _receive_output(self):
         """
-        Recieve output from token processor and store them in cache
+        Receive output from token processor and store them in cache
         """
         while True:
             try:

fastdeploy/model_executor/graph_optimization/cudagraph_piecewise_backend.py

@@ -40,7 +40,7 @@ class ConcreteSizeEntry:
     # Has runtime-bs been captured before
     captured: bool = False
 
-    # Need to be captured callable object（dynamic graph or static grpah backend）
+    # Need to be captured callable object（dynamic graph or static graph backend）
     runnable: Callable = None  # type: ignore
     # Number of completed warmups
     num_finished_warmup: int = 0

fastdeploy/model_executor/graph_optimization/graph_optimization_backend.py

@@ -117,9 +117,9 @@ class GraphOptBackend:
 
         self.max_captre_batch = fd_config.graph_opt_config.cudagraph_capture_sizes[0]
         if self.fd_config.graph_opt_config.graph_opt_level > 0:
-            # 1. Prepare cuda grpah input buffers (contain output of subgraphs)
+            # 1. Prepare cuda graph input buffers (contain output of subgraphs)
 
-            # 2. Convert dynamic grpah to static graph
+            # 2. Convert dynamic graph to static graph
 
             backend = (
                 ToStaticBackend.CINN if self.fd_config.graph_opt_config.graph_opt_level > 1 else ToStaticBackend.PHI

fastdeploy/model_executor/layers/attention/append_attn_backend.py

@@ -193,7 +193,7 @@ class AppendAttentionBackend(AttentionBackend):
         kv_cache_quant_type: str = None,
     ):
         """
-        Caculate kv cache shape
+        Calculate kv cache shape
         """
         if kv_cache_quant_type is not None and kv_cache_quant_type == "int4_zp":
             return (

fastdeploy/model_executor/layers/attention/block_multihead_attn_backend.py

@@ -114,7 +114,7 @@ class BlockAttentionBackend(AttentionBackend):
         kv_cache_quant_type: str = None,
     ):
         """
-        Caculate kv cache shape
+        Calculate kv cache shape
         """
         if kv_cache_quant_type is not None and kv_cache_quant_type == "int4_zp":
             return (

fastdeploy/model_executor/layers/attention/flash_attn_backend.py

@@ -176,7 +176,7 @@ class FlashAttentionBackend(AttentionBackend):
         kv_cache_quant_type: str = None,
     ):
         """
-        Caculate kv cache shape
+        Calculate kv cache shape
         """
         if kv_cache_quant_type is not None and kv_cache_quant_type == "int4_zp":
             return (

fastdeploy/model_executor/layers/attention/iluvatar_attn_backend.py

@@ -210,7 +210,7 @@ class IluvatarAttnBackend(AttentionBackend):
         kv_cache_quant_type: str = None,
     ):
         """
-        Caculate kv cache shape
+        Calculate kv cache shape
         """
         return (
             max_num_blocks,

fastdeploy/model_executor/layers/attention/xpu_attn_backend.py

@@ -130,7 +130,7 @@ class XPUAttentionBackend(AttentionBackend):
         kv_cache_quant_type: str = None,
     ) -> Tuple[int, int, int, int]:
         """
-        Caculate kv cache shape
+        Calculate kv cache shape
         """
         return (
             max_num_blocks,

fastdeploy/model_executor/layers/backends/gcu/attention/flash_attn_backend.py

@@ -170,7 +170,7 @@ class GCUFlashAttnBackend(AttentionBackend):
                 cache_len = 0
             elif self.seq_lens_decoder_list[seq_idx][0] != 0:  # decode
                 cache_len = self.seq_lens_decoder_list[seq_idx][0]
-            # else:  doesnot have req in this seq_idx
+            # else:  doesn't have req in this seq_idx
 
             if cache_len is not None:
                 lens_this_time = self.seq_lens_this_time_list[seq_idx]
@@ -212,7 +212,7 @@ class GCUFlashAttnBackend(AttentionBackend):
         kv_cache_quant_type: str = None,
     ):
         """
-        Caculate kv cache shape
+        Calculate kv cache shape
         """
         # [total_tokens, kv_num_heads, head_dim]
         return (

fastdeploy/model_executor/layers/backends/gcu/attention/mem_efficient_attn_backend.py

@@ -171,7 +171,7 @@ class GCUMemEfficientAttnBackend(AttentionBackend):
                 cache_len = 0
             elif self.seq_lens_decoder_list[seq_idx][0] != 0:  # decode
                 cache_len = self.seq_lens_decoder_list[seq_idx][0]
-            # else:  doesnot have req in this seq_idx
+            # else:  doesn't have req in this seq_idx
 
             if cache_len is not None:
                 lens_this_time = self.seq_lens_this_time_list[seq_idx]
@@ -224,7 +224,7 @@ class GCUMemEfficientAttnBackend(AttentionBackend):
         kv_cache_quant_type: str = None,
     ):
         """
-        Caculate kv cache shape
+        Calculate kv cache shape
         """
         # [total_tokens, kv_num_heads, head_dim]
         return (

fastdeploy/model_executor/layers/backends/metax/attention/flash_attn_backend.py

@@ -137,7 +137,7 @@ class FlashAttentionBackend(AttentionBackend):
         kv_cache_quant_type: str = None,
     ):
         """
-        Caculate kv cache shape
+        Calculate kv cache shape
         """
         if kv_cache_quant_type is not None and kv_cache_quant_type == "int4_zp":
             return (

fastdeploy/model_executor/layers/moe/ep.py

@@ -114,7 +114,7 @@ class DeepEPEngine:
                 low_latency_mode=True,
                 num_qps_per_rank=24,
             )
-        # In disaggregated mode on mutiple nodes, we either use
+        # In disaggregated mode on multiple nodes, we either use
         # high throughput mode or low latency mode.
         else:
             if moe_phase.phase == "decode":

fastdeploy/model_executor/layers/sample/early_stopper.py

@@ -35,7 +35,7 @@ class EarlyStopper:
     @abstractmethod
     def process(self, probs: paddle.Tensor, next_tokens: paddle.Tensor, stop_flags: paddle.Tensor):
         """
-        processs the stopper and set the stop_flags corresponding to the batch that triggers early stop to True
+        process the stopper and set the stop_flags corresponding to the batch that triggers early stop to True
         args:
             - probs: [batch_size, vocab_size], the probs of every sample
             - next_tokens: [batch_size, 1], the token index of every chosen sample

fastdeploy/output/token_processor.py

@@ -267,7 +267,7 @@ class TokenProcessor:
             spec_logger.info(
                 f"Speculate global accept ratio(Accept draft_tokens/Generated tokens): {accept_ratio}"
                 f" total step: {self.total_step}. total output token num: {self.number_of_output_tokens}"
-                f" avarage accept len: {self.number_of_output_tokens / self.total_step}"
+                f" average accept len: {self.number_of_output_tokens / self.total_step}"
             )
 
             if self.cfg.speculative_config.method in ["mtp"]:

fastdeploy/spec_decode/base.py

@@ -72,7 +72,7 @@ class Proposer(ABC):
     @abstractmethod
     def _run_impl(self, *args, **kwargs) -> Any:
         """
-        Implemention for different method
+        Implementation for different method
         """
         raise NotImplementedError
 

fastdeploy/worker/experts_manager.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 """
 
-"""redundant expert manger."""
+"""redundant expert manager."""
 from typing import Optional, Tuple
 
 import numpy as np

fastdeploy/worker/gcu_worker.py

@@ -49,7 +49,7 @@ class GcuWorker(WorkerBase):
     def init_device(self):
         """Initialize device and Construct model runner"""
         if paddle.is_compiled_with_custom_device("gcu"):
-            # Set evironment variable
+            # Set environment variable
             self.device_ids = self.parallel_config.device_ids.split(",")
             self.device = f"gcu:{self.local_rank}"
             paddle.device.set_device(self.device)
@@ -127,7 +127,7 @@ class GcuWorker(WorkerBase):
         # NOTE(gongshaotian): may be not need warm_up at this place
         if self.model_runner.graph_opt_level >= 1:
             self.model_runner.sot_warmup()
-        # 2. Triger cuda grpah capture
+        # 2. Trigger cuda graph capture
         self.model_runner.capture_model()
         set_random_seed(self.fd_config.model_config.seed)
 

fastdeploy/worker/gpu_worker.py

@@ -60,7 +60,7 @@ class GpuWorker(WorkerBase):
         """
         self.max_chips_per_node = 16 if current_platform.is_iluvatar() else 8
         if self.device_config.device_type == "cuda" and paddle.device.is_compiled_with_cuda():
-            # Set evironment variable
+            # Set environment variable
             self.device_ids = self.parallel_config.device_ids.split(",")
             self.device = f"gpu:{self.local_rank % self.max_chips_per_node}"
             paddle.device.set_device(self.device)
@@ -169,7 +169,7 @@ class GpuWorker(WorkerBase):
             )
         )
 
-        return available_kv_cache_memory  # return to caculate the block num in this device
+        return available_kv_cache_memory  # return to calculate the block num in this device
 
     def load_model(self) -> None:
         """Load model"""
@@ -209,7 +209,7 @@ class GpuWorker(WorkerBase):
         """
         if self.model_runner.graph_opt_level >= 1:
             self.model_runner.sot_warmup()
-        # Triger cuda grpah capture
+        # Trigger cuda graph capture
         self.model_runner.capture_model()
 
     def check_health(self) -> bool:

fastdeploy/worker/iluvatar_worker.py

@@ -51,7 +51,7 @@ class IluvatarWorker(GpuWorker):
         Initialize device and construct model runner
         """
         if paddle.is_compiled_with_custom_device("iluvatar_gpu"):
-            # Set evironment variable
+            # Set environment variable
             self.device = f"iluvatar_gpu:{self.local_rank}"
             paddle.device.set_device(self.device)
             paddle.set_default_dtype(self.parallel_config.dtype)

fastdeploy/worker/metax_worker.py

@@ -54,7 +54,7 @@ class MetaxWorker(WorkerBase):
         """
         self.max_chips_per_node = 8
         if paddle.is_compiled_with_custom_device("metax_gpu"):
-            # Set evironment variable
+            # Set environment variable
             self.device_ids = self.parallel_config.device_ids.split(",")
             self.device = f"metax_gpu:{self.local_rank % self.max_chips_per_node}"
             paddle.device.set_device(self.device)
@@ -202,7 +202,7 @@ class MetaxWorker(WorkerBase):
         """
         if self.model_runner.graph_opt_level >= 1:
             self.model_runner.sot_warmup()
-        # Todo Triger cuda grpah capture.
+        # Todo Trigger cuda graph capture.
 
     def check_health(self) -> bool:
         """ """

fastdeploy/worker/utils.py

@@ -21,7 +21,7 @@ import traceback
 def check_safetensors_model(model_dir: str):
     """
     model_dir : the directory of the model
-    Check whther the model is safetensors format
+    Check whether the model is safetensors format
     """
     model_files = list()
     all_files = os.listdir(model_dir)

fastdeploy/worker/worker_base.py

@@ -27,7 +27,7 @@ from fastdeploy.worker.output import ModelRunnerOutput
 class WorkerBase(ABC):
     """
     Engine -> (WIP)Executor -> Worker -> ModelRunner -> Model
-    Worker interface that allows inference framwork to cleanly separate implementations for different harware.
+    Worker interface that allows inference framework to cleanly separate implementations for different hardware.
     """
 
     def __init__(
@@ -89,7 +89,7 @@ class WorkerBase(ABC):
 
     @abstractmethod
     def graph_optimize_and_warm_up_model(self) -> None:
-        """Prepare model for execution through grpah optimizaiton(CudaGrpah/CINN) or warmup."""
+        """Prepare model for execution through graph optimizaiton(CudaGrpah/CINN) or warmup."""
         raise NotImplementedError
 
     @abstractmethod

fastdeploy/worker/worker_process.py

@@ -249,7 +249,7 @@ class PaddleDisWorkerProc:
         )
 
     def event_loop_normal(self) -> None:
-        """Main event loop for Paddle Distrubuted Workers.
+        """Main event loop for Paddle Distributed Workers.
         TODO(gongshaotian): support remote calling of functions that control worker.
         """
         # Currently, only support single node
@@ -493,7 +493,7 @@ def parse_args():
         "--speculative_config",
         type=json.loads,
         default=None,
-        help="Configation of SpeculativeConfig.",
+        help="Configuration of SpeculativeConfig.",
     )
     parser.add_argument(
         "--max_num_batched_tokens",
@@ -542,7 +542,7 @@ def parse_args():
         "--quantization",
         type=str,
         default="None",
-        help="Quantization name for the model, currentlly support "
+        help="Quantization name for the model, currently support "
         "'wint4', 'wint8',"
         "default is None. The priority of this configuration "
         "is lower than that of the config file. "
@@ -552,7 +552,7 @@ def parse_args():
         "--graph_optimization_config",
         type=json.loads,
         default=None,
-        help="Configation of Graph optimization backend.",
+        help="Configuration of Graph optimization backend.",
     )
     parser.add_argument(
         "--moba_attention_config",

fastdeploy/worker/xpu_worker.py

@@ -50,7 +50,7 @@ class XpuWorker(WorkerBase):
     def init_device(self):
         """Initialize device and Construct model runner"""
         if paddle.is_compiled_with_xpu():
-            # Set evironment variable
+            # Set environment variable
             self.device = f"xpu:{self.local_rank}"
             paddle.device.set_device(self.device)
             paddle.set_default_dtype(self.parallel_config.dtype)