Skip to main content

Documentation Index

Fetch the complete documentation index at: https://sambanova-systems.mintlify.dev/docs/llms.txt

Use this file to discover all available pages before exploring further.

Release notes for SambaNova Runtime, including new features, enhancements, and fixes.
For release notes older than those listed here, see the Runtime release notes archive.

SambaNova Runtime v1.26.1-2 release

Release Date: May 1, 2026 This release includes runtime performance improvements, expanded hardware support for 4-channel DDR configurations, larger model bundle support, automated SNFM database migration, and fan fault detection.

New features and enhancements

Argins performance improvements for TTFT, throughput, and model switching Reduces TTFT, token throughput, and model switching latency through three runtime optimizations:
  • Argin table structures are precomputed at model load time rather than recomputed per request
  • Schedule updates run in parallel per RDU
  • Argument tables use contiguous memory allocation to reduce DMA operations
Per-section AGCU/SLB tracking for larger model bundles Enables larger model bundles, including Llama 4 Maverick, by tracking AGCU and SLB resource usage per section rather than per segment, allowing accurate resource allocation when loading models into memory. Pipelined AGCU control commands Two separate graph control commands such as PL and AL+EXEC can now be launched together instead of sequentially, reducing CSR write time. Use CMD_PIPELINE_OPT=0 to disable the feature (enabled by default). SNFM DB migration during SND initialization SNFM automatically migrates the error and fault history databases during SND initialization when the sambanova-runtime package is installed or upgraded. Error and fault history is preserved across runtime version upgrades that change the database schema. Runtime support for 4-channel DDR configurations Adds runtime support for SambaRack SN40L-16 nodes with 4-channel (1DPC) DDR configurations:
  • Corrects DDR physical address calculation for 4-DIMM 1DPC systems
  • Updates the inventory health check so RDUs with 4 DDR channels are reported as healthy in snfadm inventory rather than degraded
Fan fault detection and reporting Runtime now detects and reports fan speed critical faults forwarded from system firmware. Fan faults are surfaced as ETYPE_FAN_SPEED_CRIT in snfadm error logs, enabling faster identification of thermal and cooling issues that could affect system stability. Optimized checkpoint loading with AVX instructions Improved performance for checkpoint loading through optimized RM→CVRM conversions using AVX instructions.

Changes

Updated default Huge Page allocation The default Huge Page allocation has been increased from ¼ to ⅓ of physical host memory, with a maximum allocation of 512 × 1 GB pages.
  • If a system has fewer Huge Pages than the new default, installing the sambanova-runtime package automatically increases the allocation.
  • If a system has more Huge Pages than the new default and you want to apply the new default, run /opt/sambaflow/bin/snconfig set hugepages with no options.

Bug fixes

  • Performance counter return values - Fixes incorrect performance counter return values.
  • Host memory huge page fix in argtable coallocate path - Fixes argtable collocation to correctly account for host memory huge page size.