EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[6/N] MoE Refactor: Cleanup MoE-related configs (#8849)

Browse Source
This commit is contained in:
Cheng Wan
2025-08-14 21:14:53 -07:00
committed by GitHub
parent 584e1ab2d0
commit 295895120d
69 changed files with 956 additions and 1037 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
test/srt/quant/test_block_int8.py
View File

@@ -5,7 +5,7 @@ import torch
from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_moe
from sglang.srt.layers.moe.topk import select_experts
from sglang.srt.layers.moe.topk import TopKConfig, select_experts
from sglang.test.test_utils import CustomTestCase
@@ -175,10 +175,13 @@ class TestW8A8BlockINT8FusedMoE(CustomTestCase):
topk_output = select_experts(
hidden_states=a,
router_logits=score,
top_k=topk,
topk_config=TopKConfig(top_k=topk, renormalize=False),
)
with torch.inference_mode():
ref_out = torch_w8a8_block_int8_moe(
a, w1, w2, w1_s, w2_s, score, topk, block_size
)
out = fused_moe(
a,
w1,
@@ -189,9 +192,6 @@ class TestW8A8BlockINT8FusedMoE(CustomTestCase):
w2_scale=w2_s,
block_shape=block_size,
)
ref_out = torch_w8a8_block_int8_moe(
a, w1, w2, w1_s, w2_s, score, topk, block_size
)
self.assertTrue(
torch.mean(torch.abs(out.to(torch.float32) - ref_out.to(torch.float32)))

4
test/srt/quant/test_int8_kernel.py
View File

@@ -5,7 +5,7 @@ import torch
from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_moe
from sglang.srt.layers.moe.topk import select_experts
from sglang.srt.layers.moe.topk import TopKConfig, select_experts
from sglang.srt.layers.quantization.int8_kernel import per_token_quant_int8
from sglang.test.test_utils import CustomTestCase
@@ -118,7 +118,7 @@ class TestW8A8Int8FusedMoE(CustomTestCase):
topk_output = select_experts(
hidden_states=a,
router_logits=score,
top_k=topk,
topk_config=TopKConfig(top_k=topk, renormalize=False),
)
out = fused_moe(
a,