[ROCm] Add AITER RoPE + KV cache fusion for MLA prefill and decode#38313

Open

khairulkabir1661 wants to merge 20 commits intovllm-project:mainfrom

khairulkabir1661:rocm-mla-aiter-rope-kv-fusion

Commits on Mar 27, 2026

Add AMD AITER MLA fusion optimization for DeepSeek models

khairulkabir1661
and
claude
committed
Add comprehensive tests for MLA fusion on AMD AITER

khairulkabir1661
and
claude
committed
Fix pre-commit issues in test_mla_fusion.py
khairulkabir1661
committed
Fix pytest mark warnings in test_mla_fusion.py
khairulkabir1661
committed
test: Remove placeholder tests from MLA fusion test suite
khairulkabir1661
committed
Fix code review issues: improve exception handling and add logging

khairulkabir1661
and
claude
committed
Fix MLA fusion: use custom op pattern and clean up tests
khairulkabir1661
committed
Fix MLA fusion tests: compare FP8-fused vs FP8-unfused
khairulkabir1661
committed
Remove test_deterministic_outputs from MLA fusion tests
khairulkabir1661
committed
Fix MLA fusion custom op registration and optimize tests
khairulkabir1661
committed
[ROCm][FP8] Add x_scale parameter support for MLA fusion (Option 2)
khairulkabir1661
committed
Fix mypy signature compatibility for x_scale/input_scale parameters
khairulkabir1661
committed
Clean up mla.py comments (lines 259-287)
khairulkabir1661
committed
Clarify q_c_scale comment in mla.py (line 240)
khairulkabir1661
committed
Clean up fusion init comments (lines 213-231)
khairulkabir1661
committed
Clean up AITER fusion helper comments (lines 15-112)
khairulkabir1661
committed
Remove test_mla_fusion.py test file
khairulkabir1661
committed
Clean up comments in fp8_utils.py
khairulkabir1661
committed
[ROCm] Add AITER RoPE + KV cache fusion for MLA prefill and decode
khairulkabir1661
committed

Commits on Mar 30, 2026

Remove @torch_compile_guard to make fusion transparent to compiler
khairulkabir1661
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[ROCm] Add AITER RoPE + KV cache fusion for MLA prefill and decode#38313

[ROCm] Add AITER RoPE + KV cache fusion for MLA prefill and decode#38313
khairulkabir1661 wants to merge 20 commits intovllm-project:mainfrom
khairulkabir1661:rocm-mla-aiter-rope-kv-fusion

Commits on Mar 27, 2026

Add AMD AITER MLA fusion optimization for DeepSeek models

Add comprehensive tests for MLA fusion on AMD AITER

Fix pre-commit issues in test_mla_fusion.py

Fix pytest mark warnings in test_mla_fusion.py

test: Remove placeholder tests from MLA fusion test suite

Fix code review issues: improve exception handling and add logging

Fix MLA fusion: use custom op pattern and clean up tests

Fix MLA fusion tests: compare FP8-fused vs FP8-unfused

Remove test_deterministic_outputs from MLA fusion tests

Fix MLA fusion custom op registration and optimize tests

[ROCm][FP8] Add x_scale parameter support for MLA fusion (Option 2)

Fix mypy signature compatibility for x_scale/input_scale parameters

Clean up mla.py comments (lines 259-287)

Clarify q_c_scale comment in mla.py (line 240)

Clean up fusion init comments (lines 213-231)

Clean up AITER fusion helper comments (lines 15-112)

Remove test_mla_fusion.py test file

Clean up comments in fp8_utils.py

[ROCm] Add AITER RoPE + KV cache fusion for MLA prefill and decode

Commits on Mar 30, 2026

Remove @torch_compile_guard to make fusion transparent to compiler