Furiosa SDK Release 2026.5.0b3

Furiosa SDK Release 2026.5.0b3#

Furiosa SDK 2026.5.0b3 is the fourth beta release of 2026.5.0. The changes since 2026.5.0b2 are summarized below.

Added video input for the Qwen3VLForConditionalGeneration, Qwen3_5ForConditionalGeneration, and Qwen3_5MoeForConditionalGeneration architectures through the server and Python APIs. Decoding video URLs requires furiosa-llm[video] and FFmpeg shared libraries.
Added video-frame sampling controls and preprocessing caching, including reuse of previously submitted videos through UUID references.

Reduced SSM decode-state memory use with eager scheduling, allowing more requests to run concurrently when state capacity limits batching.
Enabled overlap scheduling for speculative decoding with an independent draft model and a text-only target model. Embedded MTP still requires eager scheduling.
Reduced host overhead in hierarchical KV-cache management.
Refactored prefill/decode-disaggregated transfer management and host-cache staging-memory management.

Added reasoning-effort control through reasoning_effort in Chat Completions and reasoning.effort in the Responses API. Support and accepted values depend on the model’s chat template.
Added support for parallel_tool_calls=False in Chat Completions and the Responses API to return at most one tool call. For automatically parsed tool calls, token usage can still include additional generated calls omitted from the response.
Improved NPU error diagnostics with the failing kernel launch, bucket, and pipeline stage when available, and suppressed repeated details for the same underlying fault.

Fixed sliding-window prefix-cache corruption after a disaggregated worker switches from decoding to prefilling, which could cause incorrect output or NPU-sampling failures.
Fixed KV-cache sizing for data-parallel replicas sharing a single chip.
Fixed hybrid-attention cache sizing with EXPECTED_AVERAGE_SEQ_LENGTH to account for sliding-window reservations, preventing requests from stalling and triggering engine shutdown under affected configurations.
Fixed waiting-request metrics while the overlap scheduler is idle, including prefill workers in disaggregated serving.

Fixed NPU out-of-memory failures caused by runtime I/O fragmentation during high-concurrency serving with overlap scheduling and NPU sampling.
Fixed K-EXAONE startup failures with two pipeline stages across eight chips by reducing shared-memory use for runtime I/O.
Reduced excess NPU memory allocation for kernel weights in multi-chip models.
Reduced redundant synchronization when waiting for DMA transfers to complete.

Added an explicit startup error for combining prefill/decode disaggregation with L3 caching. Unset L3_ENABLED or remove the KV-transfer configuration to use a supported configuration.
Updated Rust TLS certificate-validation and trust-store dependencies in response to the eMudhra intermediate CA compromise.