minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[AMD] Fix DeepSeek-V4-Flash-FP8 on MI300 (#27972)

Browse Source
This commit is contained in:
Xinyi Song
2026-06-12 01:51:32 -07:00
committed by GitHub
parent 694cea8656
commit 371b96e210
2 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/dsv4/unified_kv_kernels/paged_prefill.py
+3 -1
View File
@@ -46,12 +46,14 @@ import torch
import triton import triton
import triton.language as tl import triton.language as tl
from sglang.srt.utils.common import is_gfx95_supported
# OPUS gfx950 paged-prefill kernel is preferred when importable; otherwise fall # OPUS gfx950 paged-prefill kernel is preferred when importable; otherwise fall
# back to the Triton implementation below. # back to the Triton implementation below.
try: try:
from aiter.ops.pa_sparse_prefill_opus import pa_sparse_prefill_opus from aiter.ops.pa_sparse_prefill_opus import pa_sparse_prefill_opus
_HAS_OPUS = True _HAS_OPUS = is_gfx95_supported()
except ImportError: except ImportError:
pa_sparse_prefill_opus = None pa_sparse_prefill_opus = None
_HAS_OPUS = False _HAS_OPUS = False
python/sglang/srt/mem_cache/memory_pool.py
+1 -1
View File
@@ -863,7 +863,7 @@ class KVCache(abc.ABC):
self.page_size = page_size self.page_size = page_size
self.dtype = dtype self.dtype = dtype
self.device = device self.device = device
if dtype in (torch.float8_e5m2, torch.float8_e4m3fn): if dtype in (torch.float8_e5m2, torch.float8_e4m3fn, torch.float8_e4m3fnuz):
# NOTE: Store as torch.uint8 because Tensor.index_put is not implemented for torch.float8_e5m2 # NOTE: Store as torch.uint8 because Tensor.index_put is not implemented for torch.float8_e5m2
self.store_dtype = torch.uint8 self.store_dtype = torch.uint8
else: else: