@@ -23,169 +23,154 @@ import torch.distributed as dist
|
||||
from vllm.logger import logger
|
||||
|
||||
from vllm_ascend.ascend_config import get_ascend_config
|
||||
from vllm_ascend.eplb.adaptor.abstract_adaptor import EplbAdaptor
|
||||
from vllm_ascend.quantization.quant_type import QuantType
|
||||
|
||||
EPLB_EXPERT_WEIGHT_NAMES = {
|
||||
(QuantType.NONE, False): ("w13_weight", "w2_weight"),
|
||||
(QuantType.NONE, True): ("w13_weight_list", "w2_weight_list"),
|
||||
(QuantType.W8A8, False): (
|
||||
"w13_weight_list",
|
||||
"w2_weight_list",
|
||||
"w13_weight_scale_fp32_list",
|
||||
"w2_weight_scale_list",
|
||||
),
|
||||
(QuantType.W8A8, True): (
|
||||
"w13_weight_list",
|
||||
"w2_weight_list",
|
||||
"w13_weight_scale_fp32_list",
|
||||
"w2_weight_scale_list",
|
||||
"fused_w1_scale_list",
|
||||
"fused_w2_scale_list",
|
||||
),
|
||||
(QuantType.W4A8, True): (
|
||||
"w13_weight_list",
|
||||
"w2_weight_list",
|
||||
"w13_weight_scale_list",
|
||||
"w2_weight_scale_list",
|
||||
"w13_scale_bias_list",
|
||||
"w2_scale_bias_list",
|
||||
),
|
||||
(QuantType.MXFP4, False): ("w13_weight", "w2_weight", "w13_weight_scale", "w2_weight_scale"),
|
||||
(QuantType.MXFP4, True): ("w13_weight", "w2_weight", "w13_weight_scale", "w2_weight_scale"),
|
||||
(QuantType.MXFP8, False): ("w13_weight", "w2_weight", "w13_weight_scale", "w2_weight_scale"),
|
||||
(QuantType.MXFP8, True): ("w13_weight", "w2_weight", "w13_weight_scale", "w2_weight_scale"),
|
||||
}
|
||||
|
||||
|
||||
class VllmEplbAdaptor(EplbAdaptor):
|
||||
class VllmEplbAdaptor:
|
||||
_registered_moe_layers: list["torch.nn.Module"] = []
|
||||
|
||||
@staticmethod
|
||||
def register_layer(layer: "torch.nn.Module") -> None:
|
||||
"""Register a MoE layer for EPLB. Called during layer initialization.
|
||||
|
||||
Only real layers call this; PPMissingLayer won't, so the registry
|
||||
naturally contains only layers on this PP rank.
|
||||
"""
|
||||
VllmEplbAdaptor._registered_moe_layers.append(layer)
|
||||
|
||||
def __init__(self, model, **args):
|
||||
super().__init__(**args)
|
||||
self.model = model
|
||||
if hasattr(model, "language_model"):
|
||||
self.model = model.language_model
|
||||
self.config = model.config.text_config
|
||||
else:
|
||||
self.model = model
|
||||
self.config = model.config
|
||||
self.rank_id = dist.get_rank()
|
||||
self.world_size = dist.get_world_size()
|
||||
self.param_dict = dict(self.model.named_parameters())
|
||||
if self.model.config.model_type == "qwen3_moe":
|
||||
self.num_dense_layers = 0
|
||||
self.global_expert_num = self.model.config.num_experts
|
||||
else:
|
||||
self.num_dense_layers = self.model.config.first_k_dense_replace
|
||||
self.global_expert_num = self.model.config.n_routed_experts
|
||||
self.num_moe_layers = self.model.config.num_hidden_layers - self.num_dense_layers
|
||||
self.init_redundancy_expert = get_ascend_config(
|
||||
).init_redundancy_expert
|
||||
self.num_dense_layers = getattr(self.config, "first_k_dense_replace", 0)
|
||||
|
||||
# TODO: init self.expert_weight_names depending on different model types, only deepseek v3 w8a8 and qwen3-moe is supported here
|
||||
if self.model.quant_config is not None:
|
||||
self.expert_weight_names = [
|
||||
"w13_weight", "w2_weight", "w13_weight_scale",
|
||||
"w13_weight_offset", "w2_weight_scale", "w2_weight_offset"
|
||||
]
|
||||
else:
|
||||
self.expert_weight_names = ["w13_weight", "w2_weight"]
|
||||
self.moe_layers = VllmEplbAdaptor._registered_moe_layers
|
||||
self.num_moe_layers = len(self.moe_layers)
|
||||
|
||||
self.expert_map_per_layer = dict(
|
||||
) # reference to expert map on device for expert map update
|
||||
self.expert_map_per_layer_cpu = dict(
|
||||
) # copy of expert map on CPU to avoid device synchronize frequently
|
||||
for layer_idx in range(self.num_moe_layers):
|
||||
self.expert_map_per_layer[self.num_dense_layers + layer_idx] = \
|
||||
self.model.get_expert_map(self.num_dense_layers + layer_idx)
|
||||
self.expert_map_per_layer_cpu = dict() # copy of expert map on CPU to avoid device synchronize frequently
|
||||
|
||||
# TODO: here we set number of buffer tensor equal to number of expert in each laryer, which can be improved
|
||||
num_buffer_tensor = torch.where(
|
||||
self.expert_map_per_layer[self.num_dense_layers] != -1)[0].numel()
|
||||
self.buffer_tensor_list: list[list[Any]] = [
|
||||
[] for _ in range(num_buffer_tensor)
|
||||
]
|
||||
self.init_buffer_tensor(num_buffer_tensor)
|
||||
# Get num_local_experts from first real MoE layer
|
||||
first_layer = self.moe_layers[0]
|
||||
self.num_local_experts = first_layer.local_num_experts
|
||||
self.ep_rank = first_layer.ep_rank
|
||||
|
||||
self.expert_param_per_layer = dict()
|
||||
self.expert_weight_key_per_layer = dict()
|
||||
self.init_expert_param_per_layer()
|
||||
|
||||
self.log2phy_map_per_layer = dict()
|
||||
for layer_idx in range(self.num_moe_layers):
|
||||
self.log2phy_map_per_layer[self.num_dense_layers + layer_idx] = \
|
||||
self.model.get_log2phy_map(self.num_dense_layers + layer_idx)
|
||||
num_buffer_tensor = self.num_local_experts
|
||||
self.buffer_tensor_list: dict[Any, list[list[Any]]] = dict()
|
||||
self.init_buffer_tensor(num_buffer_tensor)
|
||||
|
||||
self.all_topk_ids = []
|
||||
self.log2phy_map_per_layer = dict()
|
||||
for local_idx, layer in enumerate(self.moe_layers):
|
||||
self.log2phy_map_per_layer[local_idx] = layer.get_log2phy_map()
|
||||
|
||||
def init_buffer_tensor(self, num_buffer_tensor):
|
||||
for name in self.expert_weight_names:
|
||||
complete_name = "model.layers." + str(
|
||||
self.num_dense_layers) + ".mlp.experts." + name
|
||||
expert_tensor = self.param_dict[complete_name].data[
|
||||
0:num_buffer_tensor]
|
||||
buffer_tensors = torch.empty_like(expert_tensor)
|
||||
buffer_tensor_shapes: dict[Any, list[torch.Size]] = dict()
|
||||
for local_idx, _ in enumerate(self.moe_layers):
|
||||
expert_weight_key = self.expert_weight_key_per_layer[local_idx]
|
||||
expert_weight_names = EPLB_EXPERT_WEIGHT_NAMES[expert_weight_key]
|
||||
expert_tensors = [self.param_dict[f"{local_idx}.{name}"][0] for name in expert_weight_names]
|
||||
expert_tensor_shapes = [tensor.shape for tensor in expert_tensors]
|
||||
if expert_weight_key in self.buffer_tensor_list:
|
||||
assert expert_tensor_shapes == buffer_tensor_shapes[expert_weight_key], (
|
||||
f"EPLB expert weight shapes mismatch for {expert_weight_key}: "
|
||||
f"expected {buffer_tensor_shapes[expert_weight_key]}, got {expert_tensor_shapes}"
|
||||
)
|
||||
continue
|
||||
buffer_tensor_shapes[expert_weight_key] = expert_tensor_shapes
|
||||
self.buffer_tensor_list[expert_weight_key] = [[] for _ in range(num_buffer_tensor)]
|
||||
for buffer_id in range(num_buffer_tensor):
|
||||
self.buffer_tensor_list[buffer_id].append(
|
||||
buffer_tensors[buffer_id])
|
||||
for expert_tensor in expert_tensors:
|
||||
buffer_tensor = torch.empty_like(expert_tensor)
|
||||
self.buffer_tensor_list[expert_weight_key][buffer_id].append(buffer_tensor)
|
||||
|
||||
def init_expert_param_per_layer(self):
|
||||
num_local_expert = self.param_dict["model.layers." + str(self.num_dense_layers) + \
|
||||
".mlp.experts." + self.expert_weight_names[0]].data.shape[0]
|
||||
for moe_layer_id in range(self.num_moe_layers):
|
||||
layer_idx = self.num_dense_layers + moe_layer_id
|
||||
self.expert_param_per_layer[layer_idx] = list()
|
||||
for local_expert_id in range(num_local_expert):
|
||||
self.expert_param_per_layer[layer_idx].append([
|
||||
self.param_dict["model.layers." + str(layer_idx) +
|
||||
".mlp.experts." +
|
||||
name].data[local_expert_id]
|
||||
for name in self.expert_weight_names
|
||||
])
|
||||
self.param_dict = dict()
|
||||
|
||||
for local_idx, layer in enumerate(self.moe_layers):
|
||||
quant_type = QuantType.NONE if self.model.quant_config is None else layer.quant_type
|
||||
expert_weight_key = (quant_type, get_ascend_config().enable_fused_mc2 == 1)
|
||||
if expert_weight_key[0] == QuantType.W4A8MXFP:
|
||||
raise RuntimeError(f"EPLB not support {quant_type}")
|
||||
if expert_weight_key not in EPLB_EXPERT_WEIGHT_NAMES:
|
||||
raise ValueError(f"EPLB not support {quant_type} with fused MC2 {expert_weight_key[1]}")
|
||||
expert_weight_names = EPLB_EXPERT_WEIGHT_NAMES[expert_weight_key]
|
||||
self.expert_weight_key_per_layer[local_idx] = expert_weight_key
|
||||
self.expert_param_per_layer[local_idx] = list()
|
||||
for name in expert_weight_names:
|
||||
param_key = f"{local_idx}.{name}"
|
||||
self.param_dict[param_key] = getattr(layer, name)
|
||||
for local_expert_id in range(self.num_local_experts):
|
||||
per_expert_param = list()
|
||||
for name in expert_weight_names:
|
||||
per_expert_param.append(self.param_dict[f"{local_idx}.{name}"][local_expert_id])
|
||||
self.expert_param_per_layer[local_idx].append(per_expert_param)
|
||||
|
||||
def get_rank_expert_workload(self) -> torch.Tensor:
|
||||
self.moe_load = self.model.get_all_moe_loads()
|
||||
loads = [layer.moe_load for layer in self.moe_layers]
|
||||
self.moe_load = torch.stack(loads, dim=0) if loads else torch.empty(0)
|
||||
return self.moe_load
|
||||
|
||||
def get_init_expert_map(self, num_moe_layers):
|
||||
expert_map = self.model.get_all_expert_map(num_moe_layers)
|
||||
if dist.is_initialized():
|
||||
world_size = dist.get_world_size()
|
||||
|
||||
gathered = torch.empty(
|
||||
(world_size, *expert_map.shape), # [W, L, E]
|
||||
dtype=expert_map.dtype,
|
||||
device=expert_map.device)
|
||||
|
||||
dist.all_gather_into_tensor(gathered, expert_map)
|
||||
all_maps = gathered.permute(1, 0, 2)
|
||||
all_expert_maps = all_maps.cpu()
|
||||
|
||||
for layer_idx in range(num_moe_layers):
|
||||
self.expert_map_per_layer_cpu[self.num_dense_layers + layer_idx] = \
|
||||
all_expert_maps[layer_idx][self.rank_id]
|
||||
|
||||
return all_expert_maps
|
||||
|
||||
def get_init_expert_map_from_file(self, num_moe_layers, expert_map_path):
|
||||
|
||||
try:
|
||||
expert_map_tensor, layers_num, ranks_num = self._expert_file_to_tensor(
|
||||
expert_map_path)
|
||||
expert_map_all = self.local2global(expert_map_tensor)
|
||||
except (TypeError, FileNotFoundError, OSError):
|
||||
expert_map_all = self.determine_expert_map_all()
|
||||
|
||||
for layer_idx in range(num_moe_layers):
|
||||
if self.model.config.model_type == "qwen3_moe":
|
||||
self.expert_map_per_layer_cpu[layer_idx] = \
|
||||
expert_map_all[layer_idx][self.rank_id]
|
||||
else:
|
||||
self.expert_map_per_layer_cpu[layer_idx + self.num_dense_layers] = \
|
||||
expert_map_all[layer_idx][self.rank_id]
|
||||
return expert_map_all
|
||||
|
||||
def _expert_file_to_tensor(self, expert_map_path: str):
|
||||
with open(expert_map_path, "r") as f:
|
||||
data = json.load(f)
|
||||
layers_num = data["moe_layer_count"]
|
||||
gpus_num = data["layer_list"][0]["device_count"]
|
||||
|
||||
tensor_data = []
|
||||
for layer in data["layer_list"]:
|
||||
device_data = []
|
||||
for device in layer["device_list"]:
|
||||
device_data.append(device["device_expert"])
|
||||
tensor_data.append(device_data)
|
||||
expert_map_tensor = torch.tensor(tensor_data, dtype=torch.int32)
|
||||
return expert_map_tensor, layers_num, gpus_num
|
||||
logger.error(f"failed to read expert_map_path: {expert_map_path}")
|
||||
def clear_all_moe_loads(self):
|
||||
for layer in self.moe_layers:
|
||||
layer.clear_moe_load()
|
||||
|
||||
def _export_tensor_to_file(self, expert_maps, expert_map_record_path: str):
|
||||
if self.rank_id == 0:
|
||||
num_local_experts = expert_maps.max() + 1
|
||||
expert_maps_local = self.global2local(expert_maps,
|
||||
num_local_experts)
|
||||
|
||||
expert_maps_list = expert_maps_local.tolist()
|
||||
record: dict[str, Any] = {
|
||||
"moe_layer_count": len(expert_maps_list),
|
||||
"layer_list": []
|
||||
}
|
||||
expert_maps_list = expert_maps.tolist()
|
||||
record: dict[str, Any] = {"moe_layer_count": len(expert_maps_list), "layer_list": []}
|
||||
|
||||
for layer_idx, layer_data in enumerate(expert_maps_list):
|
||||
layer_record: dict[str, Any] = {
|
||||
"layer_id": layer_idx,
|
||||
"device_count": len(layer_data),
|
||||
"device_list": []
|
||||
"device_list": [],
|
||||
}
|
||||
|
||||
for device_idx, experts in enumerate(layer_data):
|
||||
device_record = {
|
||||
"device_id": device_idx,
|
||||
"device_expert": experts
|
||||
}
|
||||
placement = [experts.index(i) for i in range(num_local_experts)]
|
||||
device_record = {"device_id": device_idx, "device_expert": placement}
|
||||
layer_record["device_list"].append(device_record)
|
||||
|
||||
record["layer_list"].append(layer_record)
|
||||
@@ -194,96 +179,26 @@ class VllmEplbAdaptor(EplbAdaptor):
|
||||
json.dump(record, f, indent=4)
|
||||
|
||||
def do_update_expert_map(self, layer_id, updated_expert_map):
|
||||
self.expert_map_per_layer[layer_id] = updated_expert_map.clone()
|
||||
self.expert_map_per_layer_cpu[layer_id] = updated_expert_map.clone()
|
||||
self.expert_map_per_layer_cpu[layer_id].copy_(updated_expert_map)
|
||||
|
||||
def do_update_expert_weight(self, layer_id, local_expert_to_replace,
|
||||
buffer_tensor_id):
|
||||
def do_update_expert_weight(self, layer_id, local_expert_to_replace, buffer_tensor_id):
|
||||
expert_weight_key = self.expert_weight_key_per_layer[layer_id]
|
||||
for expert_tensor, buffer_tensor in zip(
|
||||
self.expert_param_per_layer[layer_id][local_expert_to_replace],
|
||||
self.buffer_tensor_list[buffer_tensor_id]):
|
||||
expert_tensor = buffer_tensor.clone()
|
||||
logger.debug(f"Expert tensor shape is :{expert_tensor.shape}")
|
||||
self.expert_param_per_layer[layer_id][local_expert_to_replace],
|
||||
self.buffer_tensor_list[expert_weight_key][buffer_tensor_id],
|
||||
):
|
||||
expert_tensor.copy_(buffer_tensor)
|
||||
logger.debug("Expert tensor shape is :%s", expert_tensor.shape)
|
||||
|
||||
def do_update_log2phy_map(self, layer_id, updated_log2phy_map):
|
||||
if self.log2phy_map_per_layer[layer_id] is not None:
|
||||
self.log2phy_map_per_layer[layer_id].copy_(updated_log2phy_map)
|
||||
|
||||
def global2local(self, placement: torch.Tensor,
|
||||
E_local: int) -> torch.Tensor:
|
||||
def get_global_expert_map(self):
|
||||
all_layer_global_expert_map = []
|
||||
for local_idx, layer in enumerate(self.moe_layers):
|
||||
map_cpu = layer.global_expert_map.cpu()
|
||||
all_layer_global_expert_map.append(map_cpu)
|
||||
self.expert_map_per_layer_cpu[local_idx] = map_cpu[self.ep_rank]
|
||||
|
||||
L, G, _ = placement.shape
|
||||
device = placement.device
|
||||
|
||||
pt_local = torch.full((L, G, E_local),
|
||||
fill_value=-1,
|
||||
dtype=torch.long,
|
||||
device=device)
|
||||
|
||||
valid = placement >= 0
|
||||
l_idx, g_idx, k_idx = valid.nonzero(as_tuple=True)
|
||||
|
||||
slot_idx = placement[l_idx, g_idx, k_idx]
|
||||
|
||||
pt_local[l_idx, g_idx, slot_idx] = k_idx
|
||||
|
||||
return pt_local
|
||||
|
||||
def local2global(self, placement_local: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
L, G, E_local = placement_local.shape
|
||||
device = placement_local.device
|
||||
|
||||
max_id = torch.max(placement_local)
|
||||
E_global = (max_id + 1).item() if max_id >= 0 else 0
|
||||
|
||||
if E_global == 0:
|
||||
return torch.empty((L, G, 0), dtype=torch.long, device=device)
|
||||
|
||||
placement_global = torch.full((L, G, E_global),
|
||||
fill_value=-1,
|
||||
dtype=torch.long,
|
||||
device=device)
|
||||
|
||||
valid = placement_local >= 0
|
||||
l_idx, g_idx, slot_idx = valid.nonzero(as_tuple=True)
|
||||
gid_idx = placement_local[l_idx, g_idx, slot_idx]
|
||||
|
||||
placement_global[l_idx, g_idx, gid_idx] = slot_idx
|
||||
|
||||
return placement_global
|
||||
|
||||
def determine_expert_map_all(self):
|
||||
if self.world_size == 1:
|
||||
local_ids = torch.arange(self.global_expert_num, dtype=torch.int32)
|
||||
return local_ids.view(1, 1, -1).expand(self.num_moe_layers, 1, -1)
|
||||
|
||||
local_num_experts = self.global_expert_num // self.world_size
|
||||
|
||||
expert_map_all = torch.full(
|
||||
(self.num_moe_layers, self.world_size, self.global_expert_num),
|
||||
-1,
|
||||
dtype=torch.int32)
|
||||
|
||||
for r in range(self.world_size):
|
||||
if r < self.world_size - 1:
|
||||
start = r * local_num_experts
|
||||
end = (r + 1) * local_num_experts
|
||||
local_count = local_num_experts
|
||||
else:
|
||||
start = r * local_num_experts
|
||||
end = self.global_expert_num
|
||||
local_count = self.global_expert_num - r * local_num_experts
|
||||
|
||||
if r < self.init_redundancy_expert:
|
||||
local_count += 1
|
||||
if end < self.global_expert_num:
|
||||
end += 1
|
||||
else:
|
||||
start -= 1
|
||||
|
||||
local_ids = torch.arange(local_count, dtype=torch.int32)
|
||||
expert_map_all[:, r, start:end] = local_ids.unsqueeze(0).expand(
|
||||
self.num_moe_layers, -1)
|
||||
|
||||
return expert_map_all
|
||||
return torch.stack(all_layer_global_expert_map)
|
||||
|
||||
Reference in New Issue
Block a user