Pikaliov/DCN_custom_op
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

DCNv4 INT8 patch (Level 1): int8 storage + fp32 arithmetic for SOFIA/MERIDIAN E9

Browse Source 
- Add dcnv4_int8_cuda.cu/.h: CUDA kernel (int8 values, fp16 offsets, fp32 interp,
  requantize with value_scale/output_scale)
- Add dcnv4_int8_forward(): inference-only functional wrapper (@no_grad)
- Add DCNv4Strip.forward_int8(): module-level INT8 forward (without_pointwise=True)
- Add scripts/test_dcnv4_int8.py: correctness gate (<=1 LSB, >=99% exact)
  and informational fp16 vs int8 benchmark
- Update README: INT8 API section, updated structure tree, SOFIA/MERIDIAN context

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Pikaliov
2026-06-11 16:19:21 +03:00
parent 1b3206b6a7
commit 71862bbeca
9 changed files with 539 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
DCNv4_op/DCNv4/functions/__init__.py
View File

@@ -8,4 +8,4 @@
# from .ms_flash_deform_attn_func import FlashMSDeformAttnFunction
from .flash_deform_attn_func import FlashDeformAttnFunction
from .dcnv4_func import DCNv4Function
from .dcnv4_func import DCNv4Function, dcnv4_int8_forward