Skip to content

[jax-inference-offloading] consolidate definitions for default tensor dtype #1816

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

Open
yhtang wants to merge 6 commits into
base: main
Choose a base branch
from
Open

[jax-inference-offloading] consolidate definitions for default tensor dtype #1816

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
3ad4a53
[jax-inference-offloading] Move default dtype to proto
yhtang Dec 5, 2025
83d79db
default dtype
yhtang Dec 5, 2025
0093423
Fix proto
yhtang Dec 5, 2025
5b8ea2d
Merge branch 'main' into yhtang/jio-default-dtype
yhtang Dec 5, 2025
301cd23
remove default dtype from proto
yhtang Dec 10, 2025
740c4ff
Merge branch 'main' into yhtang/jio-default-dtype
Steboss Jan 20, 2026
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
4 changes: 2 additions & 2 deletions jax-inference-offloading/jax_inference_offloading/api/param_mapping.proto
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,7 @@ message VllmParam {
}

optional TpSharding tp_sharding = 3;
optional string dtype = 4;
optional string dtype = 4 [default = 'bfloat16'];
}

message TensorSlice {
Expand Down Expand Up @@ -82,4 +82,4 @@ message ParamMapping {

message TpModelMappingSpecs {
repeated ParamMapping mappings = 1;
}
}
8 changes: 4 additions & 4 deletions jax-inference-offloading/jax_inference_offloading/vllm/extension.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -181,7 +181,7 @@ def update_weights(self, mapping_specs: TpModelMappingSpecs):

logger.debug(f'vLLM TP rank {tp_rank} receiving {param.vllm_param.name} ...')
weight = self.transport.gather(
shape, param.vllm_param.dtype or 'bfloat16',
shape, param.vllm_param.dtype,
sharding_specs.aux_dim, sharding_specs.aux_parallelism
)
logger.debug(f'vLLM TP rank {tp_rank} received {param.vllm_param.name} shape {weight.shape}')
Expand All @@ -200,7 +200,7 @@ def update_weights(self, mapping_specs: TpModelMappingSpecs):

logger.debug(f"vLLM expecting: {param.vllm_param.name} shape {shape.tolist()} raw specs {param}")

weight = self.transport.recv(shape, param.vllm_param.dtype or 'bfloat16')
weight = self.transport.recv(shape, param.vllm_param.dtype)
self._staged_weights.append((param.vllm_param.name, weight))

# TODO: make it optional
Expand Down Expand Up @@ -229,7 +229,7 @@ def update_weights_grouped(self, mapping_specs: TpModelMappingSpecs):

param_specs.append((
shape,
param.vllm_param.dtype or 'bfloat16',
param.vllm_param.dtype,
sharding_specs.aux_dim,
sharding_specs.aux_parallelism
))
Expand Down Expand Up @@ -258,7 +258,7 @@ def update_weights_grouped(self, mapping_specs: TpModelMappingSpecs):
if sharding_specs.parallelism > 0:
shape[sharding_specs.dim] //= sharding_specs.parallelism

param_specs.append((shape, param.vllm_param.dtype or 'bfloat16'))
param_specs.append((shape, param.vllm_param.dtype))
param_names.append(param.vllm_param.name)

# Receive all weights in one grouped operation
Expand Down
Loading