xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-13 10:29:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Model] Granite-4 support loading quantized checkpoint (#22925)

Browse Source 
Signed-off-by: Chih-Chieh-Yang <7364402+cyang49@users.noreply.github.com>
This commit is contained in:
Chih-Chieh Yang 2025-08-15 14:47:56 -04:00 committed by GitHub
parent 8ad7285ea2
commit 6cd69f51bf
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 6 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/model_executor/models/granitemoehybrid.py
View File

@ -471,7 +471,10 @@ class GraniteMoeHybridModel(nn.Module):
# Mapping different experts' layout: # Mapping different experts' layout:
# from HF (input_linear, output_linear, router) # from HF (input_linear, output_linear, router)
# to vLLM (experts_w13({e}.w1, {e}.w2), experts_w3({e}.w3), gate) # to vLLM (experts_w13({e}.w1, {e}.w2), experts_w3({e}.w3), gate)
if n.endswith('.block_sparse_moe.input_linear.weight'): # The renaming and parameter loading logic is the same for weight
# and weight_scale tensors so we can reuse them without issues.
if (n.endswith('.block_sparse_moe.input_linear.weight') or
n.endswith('.block_sparse_moe.input_linear.weight_scale')):
for e in range(p.size(0)): for e in range(p.size(0)):
w1_name = n.replace( w1_name = n.replace(
'.block_sparse_moe.input_linear.weight', '.block_sparse_moe.input_linear.weight',
@ -490,7 +493,8 @@ class GraniteMoeHybridModel(nn.Module):
w3_name, w3_name,
shard_id='w3', shard_id='w3',
expert_id=e) expert_id=e)
elif n.endswith('.block_sparse_moe.output_linear.weight'): elif (n.endswith('.block_sparse_moe.output_linear.weight') or
n.endswith('.block_sparse_moe.output_linear.weight_scale')):
for e in range(p.size(0)): for e in range(p.size(0)):
w2_name = n.replace( w2_name = n.replace(
'.block_sparse_moe.output_linear.weight', '.block_sparse_moe.output_linear.weight',