Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

COMPANY · ENTITY #10855

Mamba-2

Related event timeline, sources and context from the news index.

EVENT TIMELINE

4

MODELS · 1 SOURCE · AWS Machine Learning

NVIDIA Nemotron 3.5 Super VL available in Amazon SageMaker JumpStart

NVIDIA's Nemotron 3.5 Super VL — a 120B-parameter hybrid MoE vision-language model (12B active parameters) with a C-RADIO vision encoder, 1M-token context support, and an MTP head — is now available in Amazon SageMaker JumpStart. JumpStart offers a quantized NVFP4 variant optimized for inference and a BF16 full-precision reference; the BF16 weights can be downloaded for customization, and the service shows measured inference performance and deployment options on AWS instances.

7.0

CODING · 1 SOURCE · NVIDIA Developer

NVIDIA cuTile and Jane Street optimize Mamba‑2 forward pass with three Tensor Core kernels

Jane Street implemented the Mamba‑2 state‑space sequence model using NVIDIA's CUDA Tile (cuTile) Python DSL and worked with NVIDIA to optimize three custom forward‑pass kernels—chunk_cumsum, fwd_pass_state, and compute_out—targeting Tensor Cores. Optimizations include register‑resident tiling, TF32 casts, DRAM latency hints, operand routing to Tensor Memory, TMA‑friendly memory layouts, and an autotuner for block size and occupancy per input shape.

6.0

RESEARCH · 1 SOURCE · NVIDIA Developer

NVIDIA and Dao AI Lab publish ReplaySSM to speed SSM-based hybrid-model decoding

NVIDIA and the Dao AI Lab (Princeton) released ReplaySSM (paper on arXiv and code on GitHub), a technique that caches recent SSM inputs instead of full state to halve state memory traffic and speed SSM decoding kernels by ~1.46–1.80× (batch 256 on Blackwell Ultra B300). ReplaySSM also simplifies speculative-decoding rollback (1.85–2.10× higher speculative throughput on 120B-class models) and enables mixed-precision state (three of four tested models stay within ~2 percentage points of FP32 on GPQA‑Diamond); it's integrated into TensorRT‑LLM, vLLM, and FlashInfer and runs in production in GLM‑5.3‑Flash and Kimi‑K3.

8.0

RESEARCH · 1 SOURCE · arXiv cs.AI

SMat-Attention: structured causal masks that trade off routing complexity and efficiency

This arXiv preprint introduces Structured Matrix Attention (SMat-Attention), a family of causal attention masks whose row supports have VC-dimension d, letting d=1 recover the standard causal mask and larger d enable richer subset-routing. The paper gives chunkwise forward/backward algorithms for hardware efficiency, proves subquadratic prefill work O(T^{2-3/d}+T) for a prescribed family and constant-time per-token decoding after a distant prefix with O(T^{1-1/d}) cached states in fixed-horizon streaming, and reports experiments and learned-routing extensions to Mamba-2 and Gated DeltaNet showing improved recall and comparable small-scale LM performance.

7.0