Update pre-commit hooks (#12475)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -13,7 +13,7 @@
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""Inference-only Gemma model compatible with HuggingFace weights."""
|
||||
from functools import lru_cache
|
||||
from functools import cache
|
||||
from typing import Iterable, List, Optional, Set, Tuple, Union
|
||||
|
||||
import torch
|
||||
@@ -48,7 +48,7 @@ from .utils import (is_pp_missing_parameter,
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
@lru_cache(maxsize=None)
|
||||
@cache
|
||||
def _get_gemma_act_fn(
|
||||
hidden_act: Optional[str],
|
||||
hidden_activation: Optional[str],
|
||||
|
||||
@@ -429,10 +429,10 @@ class GraniteMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
||||
for e in range(p.size(0)):
|
||||
w1_name = n.replace(
|
||||
'.block_sparse_moe.input_linear.weight',
|
||||
".block_sparse_moe.experts.%d.w1.weight" % e)
|
||||
f".block_sparse_moe.experts.{e}.w1.weight")
|
||||
w3_name = n.replace(
|
||||
'.block_sparse_moe.input_linear.weight',
|
||||
".block_sparse_moe.experts.%d.w3.weight" % e)
|
||||
f".block_sparse_moe.experts.{e}.w3.weight")
|
||||
w1_param, w3_param = p[e].chunk(2, dim=0)
|
||||
assert w1_name not in new_weights
|
||||
assert w3_name not in new_weights
|
||||
@@ -442,7 +442,7 @@ class GraniteMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
||||
for e in range(p.size(0)):
|
||||
w2_name = n.replace(
|
||||
'.block_sparse_moe.output_linear.weight',
|
||||
".block_sparse_moe.experts.%d.w2.weight" % e)
|
||||
f".block_sparse_moe.experts.{e}.w2.weight")
|
||||
w2_param = p[e]
|
||||
assert w2_name not in new_weights
|
||||
new_weights[w2_name] = w2_param
|
||||
|
||||
@@ -1365,8 +1365,8 @@ class MllamaForConditionalGeneration(nn.Module, SupportsMultiModal):
|
||||
# For 1) text-only prefill and decode, 2) image-present decode.
|
||||
if image_inputs is None:
|
||||
full_text_row_masked_out_mask = (
|
||||
attn_metadata.encoder_seq_lens_tensor != 0).reshape(-1, 1).to(
|
||||
input_ids.device)
|
||||
attn_metadata.encoder_seq_lens_tensor
|
||||
!= 0).reshape(-1, 1).to(input_ids.device)
|
||||
skip_cross_attention = max(attn_metadata.encoder_seq_lens) == 0
|
||||
|
||||
# For image-present prefill.
|
||||
|
||||
@@ -81,8 +81,8 @@ class MLPSpeculator(nn.Module):
|
||||
|
||||
if self.tie_weights:
|
||||
assert (
|
||||
self.n_predict >
|
||||
1), "You cannot tie weights between stages when only 1 exists"
|
||||
self.n_predict > 1
|
||||
), "You cannot tie weights between stages when only 1 exists"
|
||||
embedding = VocabParallelEmbedding(
|
||||
config.vocab_size,
|
||||
self.inner_dim,
|
||||
|
||||
@@ -167,8 +167,8 @@ def sparsemixer(scores, jitter_eps=0.01):
|
||||
# compute mask for sparsity
|
||||
mask_logits_threshold, max_ind = scores.max(dim=-1, keepdim=True)
|
||||
factor = scores.abs().clamp(min=mask_logits_threshold)
|
||||
mask_logits_threshold = (
|
||||
(mask_logits_threshold - scores) / factor) > (2 * jitter_eps)
|
||||
mask_logits_threshold = ((mask_logits_threshold - scores) /
|
||||
factor) > (2 * jitter_eps)
|
||||
|
||||
# apply mask
|
||||
masked_gates = scores.masked_fill(mask_logits_threshold, float("-inf"))
|
||||
@@ -192,8 +192,8 @@ def sparsemixer(scores, jitter_eps=0.01):
|
||||
mask_logits_threshold, max_ind = masked_scores.max(dim=-1,
|
||||
keepdim=True)
|
||||
factor = scores.abs().clamp(min=mask_logits_threshold)
|
||||
mask_logits_threshold = (
|
||||
(mask_logits_threshold - scores) / factor) > (2 * jitter_eps)
|
||||
mask_logits_threshold = ((mask_logits_threshold - scores) /
|
||||
factor) > (2 * jitter_eps)
|
||||
|
||||
# apply mask
|
||||
masked_gates_top2 = masked_scores.masked_fill(mask_logits_threshold,
|
||||
|
||||
@@ -462,7 +462,8 @@ class _ModelRegistry:
|
||||
|
||||
|
||||
ModelRegistry = _ModelRegistry({
|
||||
model_arch: _LazyRegisteredModel(
|
||||
model_arch:
|
||||
_LazyRegisteredModel(
|
||||
module_name=f"vllm.model_executor.models.{mod_relname}",
|
||||
class_name=cls_name,
|
||||
)
|
||||
|
||||
@@ -333,10 +333,10 @@ class ModifiedWhisperEncoder(WhisperEncoder):
|
||||
return hidden_states
|
||||
|
||||
|
||||
@MULTIMODAL_REGISTRY.register_processor(UltravoxMultiModalProcessor,
|
||||
info=UltravoxProcessingInfo,
|
||||
dummy_inputs=UltravoxDummyInputsBuilder
|
||||
)
|
||||
@MULTIMODAL_REGISTRY.register_processor(
|
||||
UltravoxMultiModalProcessor,
|
||||
info=UltravoxProcessingInfo,
|
||||
dummy_inputs=UltravoxDummyInputsBuilder)
|
||||
class UltravoxModel(nn.Module, SupportsMultiModal, SupportsPP):
|
||||
|
||||
hf_to_vllm_mapper = WeightsMapper(
|
||||
|
||||
@@ -599,9 +599,8 @@ def make_empty_intermediate_tensors_factory(keys: List[str], hidden_size: int):
|
||||
device: torch.device,
|
||||
) -> IntermediateTensors:
|
||||
return IntermediateTensors({
|
||||
key: torch.zeros((batch_size, hidden_size),
|
||||
dtype=dtype,
|
||||
device=device)
|
||||
key:
|
||||
torch.zeros((batch_size, hidden_size), dtype=dtype, device=device)
|
||||
for key in keys
|
||||
})
|
||||
|
||||
|
||||
Reference in New Issue
Block a user