remove ppu backend (#3904)

Author: yao-fengchen

lmdeploy/cli/utils.py

@@ -388,7 +388,7 @@ def calib_search_scale(parser):
         )
 
     @staticmethod
-    def device(parser, default: str = 'cuda', choices: List[str] = ['cuda', 'ascend', 'maca', 'camb', 'ppu']):
+    def device(parser, default: str = 'cuda', choices: List[str] = ['cuda', 'ascend', 'maca', 'camb']):
         """Add argument device to parser."""
 
         return parser.add_argument('--device',

lmdeploy/messages.py

@@ -385,8 +385,7 @@ def __post_init__(self):
             'invalid max_prefill_token_num'
         assert self.num_gpu_blocks >= 0, 'invalid num_gpu_blocks'
         assert self.quant_policy in (0, 4, 8), 'invalid quant_policy'
-        assert self.device_type in ['cuda', 'ascend', 'maca', 'camb',
-                                    'ppu'], (f'invalid device_type: {self.device_type}')
+        assert self.device_type in ['cuda', 'ascend', 'maca', 'camb'], (f'invalid device_type: {self.device_type}')
         assert self.block_size >= 16 and (self.block_size & (self.block_size - 1)) == 0, \
             f'block_size must be >= 16 and a power of 2, but got {self.block_size}'
         if self.quant_policy > 0 and self.device_type not in ['cuda', 'ascend']:

lmdeploy/pytorch/backends/dlinfer/ppu/__init__.py

@@ -21,9 +21,6 @@ def _get_backend():
     if device_type == 'camb':
         from .dlinfer.camb import CambOpsBackend
         return CambOpsBackend
-    if device_type == 'ppu':
-        from .dlinfer.ppu import PpuOpsBackend
-        return PpuOpsBackend
     else:
         raise RuntimeError(f'Unsupported device type: {device_type}')
 

lmdeploy/pytorch/backends/dlinfer/ppu/op_backend.py

@@ -64,14 +64,7 @@ def __init__(self, func_name: str):
         self.func_name = func_name
         self.dispatched_func = self.load_and_call
         self.device_manager.register_context_callback(self.device_callback)
-        self.device_map = {
-            'cuda': 'cuda',
-            'ascend': 'dlinfer',
-            'npu': 'dlinfer',
-            'maca': 'dlinfer',
-            'camb': 'dlinfer',
-            'ppu': 'dlinfer'
-        }
+        self.device_map = {'cuda': 'cuda', 'ascend': 'dlinfer', 'npu': 'dlinfer', 'maca': 'dlinfer', 'camb': 'dlinfer'}
 
     def device_callback(self, context: DeviceContext):
         """Device context callback."""

lmdeploy/pytorch/backends/selector.py

@@ -7,12 +7,8 @@
 ASCEND_MODULE_MAP = dict()
 MACA_MODULE_MAP = dict()
 CAMB_MODULE_MAP = dict()
-PPU_MODULE_MAP = dict()
 
-DEVICE_SPECIAL_MODULE_MAP = dict(ascend=ASCEND_MODULE_MAP,
-                                 maca=MACA_MODULE_MAP,
-                                 camb=CAMB_MODULE_MAP,
-                                 ppu=PPU_MODULE_MAP)
+DEVICE_SPECIAL_MODULE_MAP = dict(ascend=ASCEND_MODULE_MAP, maca=MACA_MODULE_MAP, camb=CAMB_MODULE_MAP)
 
 # llama
 MODULE_MAP.update({

lmdeploy/pytorch/kernels/dispatcher.py

@@ -335,7 +335,7 @@ def get_max_batch_size(device_type: str):
     Args:
         device_type (str): the type of device
     """
-    assert device_type in ['cuda', 'ascend', 'maca', 'camb', 'ppu']
+    assert device_type in ['cuda', 'ascend', 'maca', 'camb']
     if device_type == 'cuda':
         max_batch_size_map = {'a100': 256, 'a800': 256, 'h100': 512, 'h800': 512}
         import torch
@@ -352,8 +352,6 @@ def get_max_batch_size(device_type: str):
         return 256
     elif device_type == 'camb':
         return 256
-    elif device_type == 'ppu':
-        return 256
 
 
 def is_bf16_supported(device_type: str = 'cuda'):
@@ -401,7 +399,6 @@ def try_import_deeplink(device_type: str):
         'npu',
         'maca',
         'camb',
-        'ppu',
     ]
     if device_type in deeplink_device_type_list:
         try: