Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

moe router tp removed #1091

Closed
wants to merge 1 commit into from
Closed

moe router tp removed #1091

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 1 addition & 6 deletions examples/llama/convert_checkpoint.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1003,13 +1003,8 @@ def convert_hf_llama(hf_model,

moe_experts_gate_weights = get_weight(
model_params, prefix + 'block_sparse_moe.gate', dtype)
v = split(moe_experts_gate_weights,
mapping.tp_size,
mapping.tp_rank,
dim=-1)

weights.update(
get_tllm_linear_weight(v.to(torch.float32),
get_tllm_linear_weight(moe_experts_gate_weights.to(torch.float32),
tllm_prex + 'mlp.router.', None,
use_weight_only,
plugin_weight_only_quant_type, dtype,
Expand Down
13 changes: 4 additions & 9 deletions tensorrt_llm/layers/moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -230,17 +230,16 @@ def __init__(self,
if quant_mode.is_weight_only():
self.weight_dtype = trt.int8

# TODO: benchmark the router and check best TP configuration
# Since output dimension is usually low (in the order of 10s), we split on input dim for the moment
# Maybe no TP at all is even more efficient
# Since output dimension is usually low (in the order of 10s), no TP at all is more efficient
# as no allreduce required in the end
self.router = RowLinear(
hidden_size,
self.num_experts,
bias=False,
dtype=trt.
float32, # Routing is sensitive since it conditions what experts are used
tp_group=tp_group,
tp_size=tp_size,
tp_group=None,
tp_size=1,
strict_dtype=True,
)

Expand Down Expand Up @@ -293,10 +292,6 @@ def __init__(self,
def forward(self, hidden_states, finished=None, lora_layer_params=None):
assert lora_layer_params is None, "LoRA + MoE is not supported for the moment"
routing_input = cast(hidden_states, trt.float32)
if self.tp_size > 1:
routing_input = split(routing_input,
self.router.in_features,
dim=-1)[self.tp_rank]
routing = self.router(routing_input)
output = _moe_plugin(self.moe_config,
hidden_states,
Expand Down