Use new dlpack api; Formatting fixes

deepspeed/autotuning/autotuner.py

@@ -71,7 +71,7 @@ def __init__(self, args, active_resources):
                 logger.info(f"Created autotuning experiments directory: {self.exps_dir}")
             except:
                 logger.error(
-                    f"Failed to create {self.exps_dir}, please check `exps_dir` in the autotuning config file is accessible by all the nodes in the job."
+                    f"Failed to create {self.exps_dir}, please check exps_dir in the autotuning config file is accessible by all the nodes in the job."
                 )
                 exit(-1)
 
@@ -84,7 +84,7 @@ def __init__(self, args, active_resources):
                 logger.info(f"Created autotuning results directory: {self.exps_dir}")
             except:
                 logger.error(
-                    f"Failed to create {self.results_dir}, please check `results_dir` in the autotuning config file is accessible by all the nodes in the job."
+                    f"Failed to create {self.results_dir}, please check results_dir in the autotuning config file is accessible by all the nodes in the job."
                 )
                 exit(-1)
 

deepspeed/inference/engine.py

@@ -600,7 +600,7 @@ def _generate(self, *inputs, **kwargs):
                 tensor_length = input_tensor.shape[-1]
                 if tensor_length > self._config.max_out_tokens:
                     raise RuntimeError(
-                        f"Input with size {tensor_length} exceeds maximum length of {self._config.max_out_tokens}. Please increase `max_tokens` in the DeepSpeed Inference Config."
+                        f"Input with size {tensor_length} exceeds maximum length of {self._config.max_out_tokens}. Please increase max_tokens in the DeepSpeed Inference Config."
                     )
 
         return self.module.generate(*inputs, **kwargs)

deepspeed/launcher/runner.py

@@ -491,7 +491,7 @@ def main(args=None):
         args.master_addr = result.decode('utf-8').split()[0]
         if not args.master_addr:
             raise RuntimeError(
-                f"Unable to detect suitable master address via `hostname -I`, please manually specify one via --master_addr"
+                "Unable to detect suitable master address via `hostname -I`, please manually specify one via --master_addr"
             )
         logger.info(f"Using IP address of {args.master_addr} for node {first_host}")
 

deepspeed/runtime/compression/cupy.py

@@ -14,10 +14,10 @@ def __init__(self):
         pass
 
     def torch2cupy(self, tensor):
-        return cupy.fromDlpack(to_dlpack(tensor))
+        return cupy.from_dlpack(to_dlpack(tensor))
 
     def cupy2torch(self, cupy_tensor):
-        return from_dlpack(cupy_tensor.toDlpack())
+        return from_dlpack(cupy_tensor)
 
     def compress_by_chunk(self, cupy_bool_tensor, num_chunks):
         packed_sign = cupy.packbits(cupy_bool_tensor)

deepspeed/runtime/engine.py

@@ -3790,7 +3790,7 @@ def save_16bit_model(self, save_dir, save_filename="pytorch_model.bin", exclude_
             else:
                 # the model will be bogus if not consolidated so don't confuse the user by saving it
                 logger.info(
-                    f"Did not save the model {path} because `stage3_gather_16bit_weights_on_model_save` is False")
+                    f"Did not save the model {path} because stage3_gather_16bit_weights_on_model_save is False")
                 return False
         else:
             state_dict = self.module_state_dict(exclude_frozen_parameters=exclude_frozen_parameters)

deepspeed/runtime/zero/partition_parameters.py

@@ -944,8 +944,7 @@ def __init__(self,
         """
         if config is not None:
             config_dict_or_path = config
-            logger.warning(
-                f'zero.Init: the `config` argument is deprecated. Please use `config_dict_or_path` instead.')
+            logger.warning('zero.Init: the `config` argument is deprecated. Please use `config_dict_or_path` instead.')
         _ds_config = deepspeed.runtime.config.DeepSpeedConfig(config_dict_or_path,
                                                               mpu) if config_dict_or_path is not None else None
         if _ds_config is not None:

deepspeed/utils/logging.py

@@ -158,6 +158,6 @@ def should_log_le(max_log_level_str):
 
     max_log_level_str = max_log_level_str.lower()
     if max_log_level_str not in log_levels:
-        raise ValueError(f"{max_log_level_str} is not one of the `logging` levels")
+        raise ValueError(f"{max_log_level_str} is not one of the logging levels")
 
     return get_current_level() <= log_levels[max_log_level_str]

op_builder/builder.py

@@ -412,8 +412,8 @@ def cpu_arch(self):
         try:
             cpu_info = get_cpu_info()
         except Exception as e:
-            self.warning(f"{self.name} attempted to use `py-cpuinfo` but failed (exception type: {type(e)}, {e}), "
-                         "falling back to `lscpu` to get this information.")
+            self.warning(f"{self.name} attempted to use py-cpuinfo but failed (exception type: {type(e)}, {e}), "
+                         "falling back to lscpu to get this information.")
             cpu_info = self._backup_cpuinfo()
             if cpu_info is None:
                 return "-march=native"
@@ -471,8 +471,8 @@ def simd_width(self):
         try:
             cpu_info = get_cpu_info()
         except Exception as e:
-            self.warning(f"{self.name} attempted to use `py-cpuinfo` but failed (exception type: {type(e)}, {e}), "
-                         "falling back to `lscpu` to get this information.")
+            self.warning(f"{self.name} attempted to use py-cpuinfo but failed (exception type: {type(e)}, {e}), "
+                         "falling back to lscpu to get this information.")
             cpu_info = self._backup_cpuinfo()
             if cpu_info is None:
                 return '-D__SCALAR__'
@@ -642,7 +642,7 @@ def compute_capability_args(self, cross_compile_archs=None):
             if cross_compile_archs_env is not None:
                 if cross_compile_archs is not None:
                     print(
-                        f"{WARNING} env var `TORCH_CUDA_ARCH_LIST={cross_compile_archs_env}` overrides `cross_compile_archs={cross_compile_archs}`"
+                        f"{WARNING} env var TORCH_CUDA_ARCH_LIST={cross_compile_archs_env} overrides cross_compile_archs={cross_compile_archs}"
                     )
                 cross_compile_archs = cross_compile_archs_env.replace(' ', ';')
             else: