Fix/qwen3.5 init failed#230

Merged

LxYuan0420 merged 19 commits intovllm-project:mainfrom

Alex-ai-future:fix/qwen3.5-init-failed

Apr 8, 2026

Commits on Apr 8, 2026

[Metal] Fix hybrid model KV cache page size alignment

Alex-ai-future
and
qwencoder
committed
[Metal] Fix Qwen3.5 hybrid model initialization

Alex-ai-future
and
qwencoder
committed
Restore model_runner.py to upstream version

Alex-ai-future
and
qwencoder
committed
Fix test for determine_available_memory single_sequence mode

Alex-ai-future
and
qwencoder
committed
[Metal] Add unit tests for update_block_size_for_backend and improve error handling

Alex-ai-future
and
qwencoder
committed
[Metal] Improve error handling in update_block_size_for_backend

Alex-ai-future
and
qwencoder
committed
[Metal] Add MLA support to update_block_size_for_backend

Alex-ai-future
and
qwencoder
committed
[Metal] Fix lint issues in MLA support changes

Alex-ai-future
and
qwencoder
committed
[Metal] Fix ruff format issues
Alex-ai-future
committed
use new device
Alex-ai-future
committed
[Metal] Address reviewer feedback on hybrid + paged attention

Alex-ai-future
and
qwencoder
committed
[Metal] Fix import order and add test for hybrid + paged error case

Alex-ai-future
and
qwencoder
committed
[Metal] Fix inaccurate docstring and test comments
Alex-ai-future
committed
[Metal] Fix test_prefix_cache.py to use mx.device_info()

Alex-ai-future
and
qwencoder
committed
[Metal] Add warning for hybrid + paged attention block-size translation

Alex-ai-future
and
qwencoder
committed
[Metal] Use cache_config.gpu_memory_utilization instead of hardcoded 0.8

Alex-ai-future
and
qwencoder
committed
[Metal] Remove code duplication in update_block_size_for_backend

Alex-ai-future
and
qwencoder
committed
[Metal] Restore one-sequence estimate in determine_available_memory

Alex-ai-future
and
qwencoder
committed
[Test] Update tests for hybrid + paged attention warning and one-sequence estimate

Alex-ai-future
and
qwencoder
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix/qwen3.5 init failed#230

Fix/qwen3.5 init failed#230
LxYuan0420 merged 19 commits intovllm-project:mainfrom
Alex-ai-future:fix/qwen3.5-init-failed

Commits on Apr 8, 2026

[Metal] Fix hybrid model KV cache page size alignment

[Metal] Fix Qwen3.5 hybrid model initialization

Restore model_runner.py to upstream version

Fix test for determine_available_memory single_sequence mode

[Metal] Add unit tests for update_block_size_for_backend and improve error handling

[Metal] Improve error handling in update_block_size_for_backend

[Metal] Add MLA support to update_block_size_for_backend

[Metal] Fix lint issues in MLA support changes

[Metal] Fix ruff format issues

use new device

[Metal] Address reviewer feedback on hybrid + paged attention

[Metal] Fix import order and add test for hybrid + paged error case

[Metal] Fix inaccurate docstring and test comments

[Metal] Fix test_prefix_cache.py to use mx.device_info()

[Metal] Add warning for hybrid + paged attention block-size translation

[Metal] Use cache_config.gpu_memory_utilization instead of hardcoded 0.8

[Metal] Remove code duplication in update_block_size_for_backend

[Metal] Restore one-sequence estimate in determine_available_memory

[Test] Update tests for hybrid + paged attention warning and one-sequence estimate