Loading…
Topic
12 posts about Quantization. Every summary links to the original.
See the idea
Blackwell inference is a pipeline, not a chip upgrade. This explainer walks five stages — fleet validation, batching, KV caching, low-precision kernels, and speculative decoding — that turn peak FLOPs into cheap tokens.
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Newest first
Hybrid models interleave state-space model (SSM) layers with attention layers, but SSM layers can become a decoding bottleneck at serving batch sizes due to memory-bound state updates. ReplaySSM represents the SSM state as a checkpoint plus a buffer of recent inputs, separating output computation from state updates. This reduces state memory traffic by roughly half, simplifies rollback for speculative decoding by avoiding state snapshots, and supports low-precision state reads without propagating rounding error through the recurrence, enabling separate precision for output computation and state updates. Evaluated on Nemotron 3 and Qwen3.5 models from 4B to 120B parameters on H100 and B300 GPUs, ReplaySSM achieves 1.20× to 1.47× faster decode steps at batch size 256, with end-to-end gains growing with batch size. The technique is integrated into TensorRT-LLM, vLLM, SGLang, and FlashInfer, and runs in production in GLM-5.3-Flash and Kimi-K3.
Transformers now supports running GGUF quantized models locally, reusing llama.cpp’s ggml kernels for efficient inference on Apple Silicon. The GGUF format packages weights, tokenizer, and chat template in one file, with quantization levels like Q4_K_M that trade precision for memory. Users can load a checkpoint from the Hub with from_pretrained and generate text using the standard Transformers API, or serve it via an OpenAI-compatible endpoint. The initial focus is on the Qwen3.5 architecture, with benchmarks on a MacBook Pro M2 Max showing token-generation rates close to llama.cpp, including prefill time. The integration uses packed quantization kernels and generation-loop optimizations to reduce overhead, though it is currently limited to single conversations on MPS without batching support.
Edge agentic workloads require repeated tool-calling turns, growing conversation histories, and valid function calls on constrained devices rather than single-prompt responses. On one NVIDIA Jetson AGX Thor Developer Kit, TensorRT Edge-LLM ran Qwen3.6-27B in SingleStream mode and completed the benchmark’s 1,007-turn performance workload in 24 minutes and 36 seconds, versus 2 hours and 37 minutes for the llama.cpp reference run. It reported 52.33 tokens per second, 247.12 ms median time to first token, and 87.94% BFCL overall accuracy. The implementation combines NVFP4 weights and activations, FP8 KV cache, reusable prompt prefixes with recurrent-state restoration, and tree-based multi-token prediction; roughly 96% of prompt tokens used hot cache, while tree-based MTP could achieve about 40% more decoding performance than linear MTP with three draft steps.
Production-scale mixture-of-experts (MoE) training must handle uneven, data-dependent expert loads, creating ragged tensors and expensive dispatch, combine, and all-to-all communication. Dropless MoE processes every selected token, preserving model quality but requiring kernels that support variable token counts without padding or dropping. The post presents JAX and NVIDIA Transformer Engine optimizations centered on group-aware MXFP8 quantization, grouped GEMM, and fused expert-parallel (EP) dispatch and combine using NCCL EP. Grouped GEMM computes each expert’s actual workload in one kernel path, while EP reduces movement overhead and can deduplicate tokens crossing networks. For DeepSeek-V3 training on NVIDIA GB200, the reported throughput increased from 103 to 1,068 TFLOPS/GPU, a 10.4x improvement; the post also reports roughly 10x throughput improvement and 97% scaling efficiency at 1,024 GPUs.
The post evaluates whether W8A8 INT8 quantization can reduce the size of a Llama 3.1 8B Instruct model without sacrificing accuracy or inference performance. Using SmoothQuant to address activation outliers and GPTQ with error compensation for BF16 weights, the model was compressed from 14.9 GB to 8.0 GB after calibration on WikiText-2. Identical zero-shot lm-eval-harness tests across MMLU, ARC Easy, HellaSwag, and IFEval found MMLU, ARC Easy, and HellaSwag deltas smaller than their standard errors, while the IFEval change was described as likely within a similar noise range; the post concludes no measurable degradation. vLLM and GuideLLM tests on one NVIDIA L40S reported 29% more concurrent requests, 24% lower synchronous TTFT, and 44% faster token generation at maximum load, while ITL degraded more quickly under high concurrency because of dynamic activation quantization overhead.
Pinterest describes how its Manas distributed search platform is evolving embedding retrieval for a corpus growing toward tens of billions of embeddings. Traditional HNSW indexes require all data in RAM, while single-vector two-tower retrieval limits context-dependent relevance, motivating changes across quantization, SSD-based serving, and multi-embedding retrieval. Manas uses Scalar Quantization and Product Quantization across HNSW and IVF, with SQ preserving over 90% recall while reducing index size, and Linear Scaling plus SIMD cuts per-query computing resources by 10–15%. For disk-based serving, SPANN with PQ is reported as 4.5x faster than plain SPANN and over 3x DiskANN QPS, while a five-billion-embedding production evaluation saved over 40% CPU time with under 5% recall drop. The platform is also adding ColBERT-style late interaction using Sum of MaxSim, with a pilot in progress.
Serving Llama 3.1 8B in BF16 requires roughly 14.9 GB for weights, leaving little VRAM on a 16 GB GPU for KV cache and activations, which limits practical deployment. The workflow benchmarks the base model, applies INT8 W8A8 quantization with llm-compressor, then repeats accuracy and performance tests under identical hardware and load conditions. SmoothQuant modifies weights to reduce activation outliers, while GPTQ uses calibration data and error compensation before rounding weights to INT8; the experiment used 512 WikiText-2 samples and targeted deployment with vLLM. On the Llama 3.1 8B Instruct model, compression reduced storage from 14.9 GB to 8.0 GB, a 46% reduction, and accuracy was preserved across MMLU, ARC, HellaSwag, and IFeval; generation-speed results are reserved for part 2.
Reasoning and agentic models have typically required data-center inference, creating network dependencies, higher costs, and potential data exposure for edge applications. The post explains how to choose between dense Qwen3.8-27B, which activates all 27 billion parameters, and mixture-of-experts Nemotron 3.5 Lightning, which has 30 billion total parameters but activates 3 billion per token. Deployment uses vLLM on Jetson with NVFP4 quantization and speculative decoding, where a draft model proposes tokens and the main model verifies them; the examples include DSpark for Nemotron and DFlash2 for Qwen. Reported tests show up to 6.28x decode-throughput speedup over BF16, while workload-specific throughput still varies. Validation with representative prompts is recommended before deployment, and custom quantization or speculator training is reserved for cases where accuracy or acceptance rates reveal a clear gap.
NeoMME is a family of 260M and 800M multilingual multimodal encoders designed to represent text and images without a separate pretrained vision tower or causal language model. A single bidirectional Transformer processes text tokens and raw 32×32 image patches, while pretraining from scratch uses a masked discrete-diffusion text-denoising objective with high masking to encourage image-grounded reconstruction. The retrieval variant adds jointly trained dense mean-pooled and 128-dimensional late-interaction heads, returning both in one forward pass for page-image retrieval. On ViDoRe v3, the 260M model reaches 0.523 and both sizes lie on the model-size Pareto frontier; at 2048×2048, it encodes about 51 pages per second on an NVIDIA L40S, roughly twice ColModernVBERT's throughput. Hierarchical token pooling and asymmetric quantization reduce late-interaction storage from roughly 1.5 MB to 6 kB per page, a 255× reduction, while retaining more than 95% of baseline nDCG@10.
LLM quantization addresses the gap between rapidly growing model sizes and GPU memory by storing weights and activations at lower precision than BF16. It explains how quantization targets linear layers in transformer blocks, distinguishes weight quantization for memory-bound decoding from activation quantization for compute-bound prefill, and maps formats such as GGUF, FP8, NVFP4, AWQ-INT4, and MLX to their inference engines. Calibrated methods including GPTQ, AWQ, and SmoothQuant protect important weights using representative data, while reported NVFP4 results recover about 99% of BF16 accuracy for 70B–235B models and 95–98% for 7B–14B models. In a Llama 3 70B benchmark on two H100s, FP8 weights-and-activations increased throughput from 158 to 474 input tokens/sec and reduced time to first token under load from over 30,000ms to about 4,800ms.
Google introduces the LiteRT Qualcomm AI Engine Direct (QNN) Accelerator, developed with Qualcomm to replace the previous TFLite QNN delegate and simplify NPU deployment across supported Qualcomm SoCs. It integrates vendor compilers and runtimes behind a unified workflow, supports ahead-of-time and on-device compilation, and can fall back to GPU or CPU when needed. Across 72 canonical vision, audio, and NLP models, the accelerator supports 90 LiteRT ops, enables full NPU delegation for 64, and reports up to 100x CPU and 10x GPU speedups. On Snapdragon 8 Elite Gen 5, FastVLM-0.5B uses int8 weights, int16 activations, and specialized transformer kernels to reach 0.12 second time-to-first-token, over 11,000 prefill tokens per second, and over 100 decode tokens per second.
An upgrade from Hopper H100s to Blackwell B200s exposed the Mixture-of-Experts (MoE) layer, implemented with MegaBlocks, as a major training bottleneck: it consumed nearly 53% of forward-pass time and 27% of backward-pass time. The team rewrote the layer at the GPU-kernel level using CUDA, PTX, and some ThunderKittens, while moving from BF16 to microscaled FP8 (MXFP8) with nearly zero loss in training quality. The implementation addresses Blackwell-specific tensor memory constraints and quantization overhead, including custom MXFP8 quantization that sustains more than 6.2 TB/s and emits scale matrices directly compatible with tcgen05.mma. On an internal model, MoE execution became 3.5x faster in both passes, producing a 1.5x end-to-end speedup on Blackwell and 2x versus the original Hopper setup, measured in tokens per second per GPU.
The best engineering reads, once a week.
