Skip to content

Make Ray engine and worker process prefill returning first token #147

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

Merged
merged 3 commits into from
Jul 12, 2024
Merged

Make Ray engine and worker process prefill returning first token #147

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
01e489c
result_token fix
richardsliu Jul 12, 2024
34b1a81
fix return
richardsliu Jul 12, 2024
ddfa9c3
Format
richardsliu Jul 12, 2024
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 jetstream_pt/ray_engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,7 +70,7 @@ def prefill(
existing_prefix: Optional[Prefix] = None,
padded_tokens: np.ndarray, # PrefillInputs[np.ndarray],
true_length: int,
) -> Prefix:
) -> Tuple[Prefix, engine_api.ResultTokens]:
if self.is_disaggregated:
return self.prefill_impl(
params=params,
Expand All @@ -95,7 +95,7 @@ def prefill_impl(
existing_prefix: Optional[Prefix] = None,
padded_tokens: np.ndarray, # PrefillInputs[np.ndarray],
true_length: int,
) -> Prefix:
) -> Tuple[Prefix, engine_api.ResultTokens]:
all_outputs = []
for worker in self.engine_workers:
prefill_func = (
Expand Down
44 changes: 40 additions & 4 deletions jetstream_pt/ray_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -461,7 +461,7 @@ def prefill_ray(
existing_prefix: Optional[Prefix] = None,
padded_tokens: PrefillInputs, # PrefillInputs[np.ndarray],
true_length: int,
) -> None:
) -> tuple[Prefix, engine_api.ResultTokens]:
"""Do prefill in ray worker"""
logits, updated_caches = self.prefill(
params=params,
Expand All @@ -476,7 +476,25 @@ def prefill_ray(
prefix = Prefix(token, updated_caches, true_length)
self.prefix_queue.put(prefix, block=False)

return token
token_out = jnp.reshape(token, (1, 1))
data = jnp.concatenate(
[
token_out, # First token
jnp.ones_like(token_out), # validity of first token
jnp.zeros((1, 1), dtype=jnp.int32), # length = 0
],
axis=-1,
)
length = token_out.shape[1]
result = engine_api.ResultTokens(
data=data,
tokens_idx=(0, length),
valid_idx=(length, 2 * length),
length_idx=(2 * length, 2 * length + 1),
samples_per_slot=1,
)

return prefix, result

def _convert_to_np_caches(
self, caches: List[Tuple[jax.Array, jax.Array]]
Expand All @@ -495,7 +513,7 @@ def prefill_ray_disaggregation(
existing_prefix: Optional[Prefix] = None,
padded_tokens: PrefillInputs, # PrefillInputs[np.ndarray],
true_length: int,
) -> Any:
) -> tuple[NpPrefix, engine_api.ResultTokens]:
"""Do prefill in ray worker"""
logits, updated_caches = self.prefill(
params=params,
Expand All @@ -513,7 +531,25 @@ def prefill_ray_disaggregation(
np_update_caches = self._convert_to_np_caches(updated_caches)
np_prefix = NpPrefix(token, np_update_caches, true_length)

return np_prefix
token_out = jnp.reshape(token, (1, 1))
data = jnp.concatenate(
[
token_out, # First token
jnp.ones_like(token_out), # validity of first token
jnp.zeros((1, 1), dtype=jnp.int32), # length = 0
],
axis=-1,
)
length = token_out.shape[1]
result = engine_api.ResultTokens(
data=data,
tokens_idx=(0, length),
valid_idx=(length, 2 * length),
length_idx=(2 * length, 2 * length + 1),
samples_per_slot=1,
)

return np_prefix, result

def transfer(self, np_prefix: NpPrefix) -> Any:
"""Transfer prefill result from object store to HBM"""
Expand Down
Loading