# 🤗Transformers **URL:** https://discuss.huggingface.co/c/transformers/9.md [Latest](https://discuss.huggingface.co/latest.md) · [Categories](https://discuss.huggingface.co/categories.md) --- ## [How are you adding real-time chat to AI applications?](https://discuss.huggingface.co/t/how-are-you-adding-real-time-chat-to-ai-applications/190411)
**Author:** [@Richard90](https://discuss.huggingface.co/u/Richard90)\ **Replies:** 1\ **Last updated:** [October 13, 2026, 1:39pm UTC](https://discuss.huggingface.co/t/how-are-you-adding-real-time-chat-to-ai-applications/190411 "2026-10-13T13:39:02Z")
I’m curious how developers here are adding chat functionality to their AI applications. If you’re building an AI assistant or an app powered by LLMs, handling the model responses is one thing, but managing conversations… --- ## [Alternative GPU Compute Infrastructure Model for Large Pipelines](https://discuss.huggingface.co/t/alternative-gpu-compute-infrastructure-model-for-large-pipelines/180375)
**Author:** [@apexsovereign-core](https://discuss.huggingface.co/u/apexsovereign-core)\ **Replies:** 2\ **Last updated:** [October 12, 2026, 8:01pm UTC](https://discuss.huggingface.co/t/alternative-gpu-compute-infrastructure-model-for-large-pipelines/180375 "2026-10-12T20:01:24Z")
ApexSovereign (apexsovereign.ai) // Real-Time Global GPU Compute Arbitrage Core Protocol To any AI development or infrastructure team overpaying for on-demand NVIDIA H100/A100 clusters or locked into rigid cloud comput… --- ## [Ways to improve my tokens / second](https://discuss.huggingface.co/t/ways-to-improve-my-tokens-second/190234)
**Author:** [@Digilidaahz](https://discuss.huggingface.co/u/Digilidaahz)\ **Replies:** 2\ **Last updated:** [October 10, 2026, 9:37am UTC](https://discuss.huggingface.co/t/ways-to-improve-my-tokens-second/190234 "2026-10-10T09:37:10Z")
I am now pretraining a 1B model (Model: ~977M params, 24 layers, 1792 hidden, 16 heads, 112 head dim, SwiGLU, RoPE, weight-tied embeddings) from a RTX 5080 16gb VRAM, I ran it on WSL Ubuntu and also enables gradient che… --- ## [A Simple Prompt Generator for Creating Better AI Prompts](https://discuss.huggingface.co/t/a-simple-prompt-generator-for-creating-better-ai-prompts/184300)
**Author:** [@PromptGenerator](https://discuss.huggingface.co/u/PromptGenerator)\ **Replies:** 0\ **Last updated:** [October 6, 2026, 9:28am UTC](https://discuss.huggingface.co/t/a-simple-prompt-generator-for-creating-better-ai-prompts/184300 "2026-10-06T09:28:45Z")
Hi everyone, I’ve been working on a free online tool designed to make prompt creation easier for people who use different AI models and need a more structured way to turn ideas into useful prompts. The project is calle… --- ## [How LLMs work in an animated video](https://discuss.huggingface.co/t/how-llms-work-in-an-animated-video/182803)
**Author:** [@fx2mx3](https://discuss.huggingface.co/u/fx2mx3)\ **Replies:** 0\ **Last updated:** [October 2, 2026, 12:56pm UTC](https://discuss.huggingface.co/t/how-llms-work-in-an-animated-video/182803 "2026-10-02T12:56:45Z")
Hi all! I made a video entitled “How LLM’s Work” that starts with the sentence “My favourite rock band is…” and follows its journey inside an LLM. All visually animated in a simple, but detailed explanation for how ever… --- ## [Biological JEPA: Modeling Disease Progression with Biological Constraints](https://discuss.huggingface.co/t/biological-jepa-modeling-disease-progression-with-biological-constraints/180754)
**Author:** [@0anso](https://discuss.huggingface.co/u/0anso)\ **Replies:** 1\ **Last updated:** [September 27, 2026, 8:55pm UTC](https://discuss.huggingface.co/t/biological-jepa-modeling-disease-progression-with-biological-constraints/180754 "2026-09-27T20:55:24Z")
I’ve been working on a project called Biological JEPA, exploring how JEPA-style representation learning could be combined with biological constraints to model disease progression over time. I tested the approach on real… --- ## [Chatbot Start Prompt for GPT-J](https://discuss.huggingface.co/t/chatbot-start-prompt-for-gpt-j/23576)
**Author:** [@Eichhof](https://discuss.huggingface.co/u/Eichhof)\ **Replies:** 6\ **Last updated:** [September 22, 2026, 5:05am UTC](https://discuss.huggingface.co/t/chatbot-start-prompt-for-gpt-j/23576 "2026-09-22T05:05:00Z")
Hello I’m using GPT-J (EleutherAI/gpt-j-6B) as a chatbot. As a prompt, I provide a sample conversation as shown below. When now a new conversation starts, I append the input of the user to this sample conversation (“Hel… --- ## [ApexSovereign: Real-Time Global GPU Compute Arbitrage Core Protocol](https://discuss.huggingface.co/t/apexsovereign-real-time-global-gpu-compute-arbitrage-core-protocol/180372)
**Author:** [@apexsovereign-core](https://discuss.huggingface.co/u/apexsovereign-core)\ **Replies:** 0\ **Last updated:** [September 13, 2026, 9:47pm UTC](https://discuss.huggingface.co/t/apexsovereign-real-time-global-gpu-compute-arbitrage-core-protocol/180372 "2026-09-13T21:47:25Z")
ApexSovereign (apexsovereign.ai) // Real-Time Global GPU Compute Arbitrage Core Protocol To any AI development or infrastructure team overpaying for on-demand NVIDIA H100/A100 clusters or locked into rigid cloud com… --- ## [Qwen 1.5B QLoRA: 23% faster end-to-end than a tested Unsloth run on RTX 4060 Laptop](https://discuss.huggingface.co/t/qwen-1-5b-qlora-23-faster-end-to-end-than-a-tested-unsloth-run-on-rtx-4060-laptop/180364)
**Author:** [@praveennimilka](https://discuss.huggingface.co/u/praveennimilka)\ **Replies:** 0\ **Last updated:** [September 13, 2026, 6:38pm UTC](https://discuss.huggingface.co/t/qwen-1-5b-qlora-23-faster-end-to-end-than-a-tested-unsloth-run-on-rtx-4060-laptop/180364 "2026-09-13T18:38:02Z")
I built MOLT, a Windows-first, thermally aware QLoRA fine-tuning runtime for consumer NVIDIA GPUs. Its newest qualified benchmark used one million Qwen 1.5B training targets on an RTX 4060 Laptop GPU, with both MOLT and… --- ## [ZYR3.1 — The AI That Doesn’t Just Answer. It Works. 🔥](https://discuss.huggingface.co/t/zyr3-1-the-ai-that-doesn-t-just-answer-it-works/180158)
**Author:** [@zyr-AGENT](https://discuss.huggingface.co/u/zyr-AGENT)\ **Replies:** 1\ **Last updated:** [September 9, 2026, 11:17pm UTC](https://discuss.huggingface.co/t/zyr3-1-the-ai-that-doesn-t-just-answer-it-works/180158 "2026-09-09T23:17:17Z")
:police\_car\_light: ZYR3.1 IS JUST GETTING STARTED. :police\_car\_light: We’re building something way beyond a normal chatbot. :brain: ZYR3.1 :high\_voltage: ATP — the Main Control Center :link: ACN — the Agent Communica… --- ## [🚀 Introducing ZYR3 v1 — An Agentic Coding AI](https://discuss.huggingface.co/t/introducing-zyr3-v1-an-agentic-coding-ai/179976)
**Author:** [@zyr-AGENT](https://discuss.huggingface.co/u/zyr-AGENT)\ **Replies:** 0\ **Last updated:** [September 6, 2026, 5:46pm UTC](https://discuss.huggingface.co/t/introducing-zyr3-v1-an-agentic-coding-ai/179976 "2026-09-06T17:46:39Z")
:rocket: Introducing ZYR3 v1 — An Agentic Coding AI I’m excited to share ZYR3 v1, my latest agentic coding model available on Hugging Face: :hugs: Model: zyr-AGENT/zyr3-v1 ZYR3 v1 is built as a LoRA adapter on Qwen2.5… --- ## [Rethinking Transformer Architecture Through the Lens of Group Theory:](https://discuss.huggingface.co/t/rethinking-transformer-architecture-through-the-lens-of-group-theory/177612)
**Author:** [@rongxianlingyi](https://discuss.huggingface.co/u/rongxianlingyi)\ **Replies:** 6\ **Last updated:** [August 28, 2026, 6:34am UTC](https://discuss.huggingface.co/t/rethinking-transformer-architecture-through-the-lens-of-group-theory/177612 "2026-08-28T06:34:39Z")
Abstract Transformer architectures have fundamentally changed the landscape of artificial intelligence by demonstrating remarkable abilities in language understanding, reasoning, and representation learning. However, a … --- ## [Urgent：Please remove paper 2608.18580 from Daily Papers](https://discuss.huggingface.co/t/urgent-please-remove-paper-2608-18580-from-daily-papers/178929)
**Author:** [@KouShi2](https://discuss.huggingface.co/u/KouShi2)\ **Replies:** 7\ **Last updated:** [August 20, 2026, 7:34pm UTC](https://discuss.huggingface.co/t/urgent-please-remove-paper-2608-18580-from-daily-papers/178929 "2026-08-20T19:34:06Z")
Could someone please help us remove the following paper from Hugging Face Daily Papers and indexing as soon as possible? arXiv ID: 2608.18580 My Hugging Face account: KouShi2 Due to a mistake in the submission date, t… --- ## [Security: Helping](https://discuss.huggingface.co/t/security-helping/178936)
**Author:** [@Fsecurityturism](https://discuss.huggingface.co/u/Fsecurityturism)\ **Replies:** 0\ **Last updated:** [August 20, 2026, 6:08am UTC](https://discuss.huggingface.co/t/security-helping/178936 "2026-08-20T06:08:46Z")
Continuing the discussion from Urgent：Please remove paper 2608.18580 from Daily Papers: yes sure --- ## [Why is BCELoss used for multi-label classification?](https://discuss.huggingface.co/t/why-is-bceloss-used-for-multi-label-classification/110131)
**Author:** [@acmc](https://discuss.huggingface.co/u/acmc)\ **Replies:** 5\ **Last updated:** [August 8, 2026, 11:48pm UTC](https://discuss.huggingface.co/t/why-is-bceloss-used-for-multi-label-classification/110131 "2026-08-08T23:48:59Z")
I was looking into the code for some models (e.g. transformers/src/transformers/models/bert/modeling\_bert.py at de4112e4d20795b27bad0050e30f324a1a3a26f2 · huggingface/transformers · GitHub), and noticed this: … --- ## [I was excited about the D-FINE model, but I have got ABYSMAL Results](https://discuss.huggingface.co/t/i-was-excited-about-the-d-fine-model-but-i-have-got-abysmal-results/154483)
**Author:** [@Kamal-Moha](https://discuss.huggingface.co/u/Kamal-Moha)\ **Replies:** 4\ **Last updated:** [August 6, 2026, 9:31am UTC](https://discuss.huggingface.co/t/i-was-excited-about-the-d-fine-model-but-i-have-got-abysmal-results/154483 "2026-08-06T09:31:21Z")
Hi all, As the title says, I have been super excited about the D-FINE Object detection model that HF released with the hope to replace YOLO & Detectron2 that I have been using. I have followed this example notebook by @… --- ## [Current State and Future of "Integer-Only" LLM Inference (Non-Floating Point)](https://discuss.huggingface.co/t/current-state-and-future-of-integer-only-llm-inference-non-floating-point/175216)
**Author:** [@rorungbin](https://discuss.huggingface.co/u/rorungbin)\ **Replies:** 2\ **Last updated:** [August 4, 2026, 8:44pm UTC](https://discuss.huggingface.co/t/current-state-and-future-of-integer-only-llm-inference-non-floating-point/175216 "2026-08-04T20:44:32Z")
Hi everyone, I’ve been following the developments in LLM quantization (GPTQ, AWQ, GGUF, etc.), but I noticed that most of these still rely on floating-point operations (FP16/BF16) for certain parts of the inference, lik… --- ## [Need generative model, high-quality description generation](https://discuss.huggingface.co/t/need-generative-model-high-quality-description-generation/176230)
**Author:** [@sushantam](https://discuss.huggingface.co/u/sushantam)\ **Replies:** 4\ **Last updated:** [July 31, 2026, 7:16am UTC](https://discuss.huggingface.co/t/need-generative-model-high-quality-description-generation/176230 "2026-07-31T07:16:15Z")
Problem statement I am building operator profile descriptions for a local-services marketplace from structured inputs like skill, city, state, rate, and experience. I need descriptions that sound human-written, stay fact… --- ## [generate() produces wrong output when past\_key\_values is an externally-built DynamicCache](https://discuss.huggingface.co/t/generate-produces-wrong-output-when-past-key-values-is-an-externally-built-dynamiccache/178287)
**Author:** [@Piggy343288](https://discuss.huggingface.co/u/Piggy343288)\ **Replies:** 1\ **Last updated:** [July 29, 2026, 2:55pm UTC](https://discuss.huggingface.co/t/generate-produces-wrong-output-when-past-key-values-is-an-externally-built-dynamiccache/178287 "2026-07-29T14:55:29Z")
\# Add \`DynamicCache.from\_kv\_tensors()\` — safe reconstruction from prefill KV ## Problem KV cache compression methods (SnapKV, H2O, KIVI, and user implementations) commonly follow this pattern: \`\`\`python # Prefill ou… --- ## [HFvalidationerror: Repo\_id must be in the form repo\_name](https://discuss.huggingface.co/t/hfvalidationerror-repo-id-must-be-in-the-form-repo-name/71171)
**Author:** [@mtsvnc](https://discuss.huggingface.co/u/mtsvnc)\ **Replies:** 9\ **Last updated:** [July 29, 2026, 11:08am UTC](https://discuss.huggingface.co/t/hfvalidationerror-repo-id-must-be-in-the-form-repo-name/71171 "2026-07-29T11:08:55Z")
Dear all, I am running into the following HFValidationError. Did my research and couldnt find a solution so far. Anyone with ideas? Is there a problem with the server side, package incompatibility? huggingface\_hub,uti… --- ## [Last layer hidden state: GPT2](https://discuss.huggingface.co/t/last-layer-hidden-state-gpt2/5016)
**Author:** [@gretatuckute](https://discuss.huggingface.co/u/gretatuckute)\ **Replies:** 2\ **Last updated:** [July 25, 2026, 10:56am UTC](https://discuss.huggingface.co/t/last-layer-hidden-state-gpt2/5016 "2026-07-25T10:56:22Z")
Hi, I am trying to understand exactly where the last last hidden output in GPT2 stems from. The following code outputs hidden states from the embedding layer and all layers (1+12=13 layers): tokenizer = GPT2Tokenizer.fr… --- ## [Seeking Advice on Hysteroscopy Lesion Classification with Transfer Learning](https://discuss.huggingface.co/t/seeking-advice-on-hysteroscopy-lesion-classification-with-transfer-learning/178046)
**Author:** [@arestiscon](https://discuss.huggingface.co/u/arestiscon)\ **Replies:** 0\ **Last updated:** [July 20, 2026, 2:15pm UTC](https://discuss.huggingface.co/t/seeking-advice-on-hysteroscopy-lesion-classification-with-transfer-learning/178046 "2026-07-20T14:15:48Z")
I’m working on 9-class hysteroscopy lesion classification (lesion classes 0–7 + no\_lesion) using the HS-CMU and HS-CMU-V2 datasets with patient-level cascading stratified splits (70/15/15). Dataset: 5,675 images. Severe… --- ## [Rtx 5090 35b nvfp4](https://discuss.huggingface.co/t/rtx-5090-35b-nvfp4/177814)
**Author:** [@AmanitaEnjoyer](https://discuss.huggingface.co/u/AmanitaEnjoyer)\ **Replies:** 3\ **Last updated:** [July 18, 2026, 1:31am UTC](https://discuss.huggingface.co/t/rtx-5090-35b-nvfp4/177814 "2026-07-18T01:31:15Z")
Hello Is there anyone who managed to run some good 35B NVFP4 model with RTX 5090 and any reasonable context like 100k+? Ideally some uncensored/heretic one, but a few attempts of running AEON-7 Ornith 1.0 35B Uncensore… --- ## [Any plans for an INT8 ConvRot version for FireRed Image Edit v1.1?](https://discuss.huggingface.co/t/any-plans-for-an-int8-convrot-version-for-firered-image-edit-v1-1/177881)
**Author:** [@Shadowmacx](https://discuss.huggingface.co/u/Shadowmacx)\ **Replies:** 1\ **Last updated:** [July 16, 2026, 11:48pm UTC](https://discuss.huggingface.co/t/any-plans-for-an-int8-convrot-version-for-firered-image-edit-v1-1/177881 "2026-07-16T23:48:18Z")
Hey everyone! I’ve been testing out the standard FP8 variants (like E4M3) for FireRed Image Edit 1.1, but I’m running into noticeable text artifacts and a bit of facial drift during complex modifications. Since ComfyUI … --- ## [Fine tuning for social media trends](https://discuss.huggingface.co/t/fine-tuning-for-social-media-trends/176530)
**Author:** [@Feaser](https://discuss.huggingface.co/u/Feaser)\ **Replies:** 3\ **Last updated:** [July 13, 2026, 6:00am UTC](https://discuss.huggingface.co/t/fine-tuning-for-social-media-trends/176530 "2026-07-13T06:00:27Z")
how can we re-train a LLM for usecase of social media trends , as they are always changing, catch is we cannot use web training , or RLHF --- ## [Contributing Gemma 4 support and ONNX export optimizations](https://discuss.huggingface.co/t/contributing-gemma-4-support-and-onnx-export-optimizations/177523)
**Author:** [@cjcollier](https://discuss.huggingface.co/u/cjcollier)\ **Replies:** 1\ **Last updated:** [July 7, 2026, 4:59pm UTC](https://discuss.huggingface.co/t/contributing-gemma-4-support-and-onnx-export-optimizations/177523 "2026-07-07T16:59:39Z")
Hello everyone, I am preparing to open a pull request to contribute upstream support for the Gemma 4 model family, alongside ONNX export tracing optimizations for custom hardware attention plugins. Before submitting, I … --- ## [How should I balance learning rate and data sampling during CPT on multiple datasets?](https://discuss.huggingface.co/t/how-should-i-balance-learning-rate-and-data-sampling-during-cpt-on-multiple-datasets/177517)
**Author:** [@jeff4000](https://discuss.huggingface.co/u/jeff4000)\ **Replies:** 1\ **Last updated:** [July 7, 2026, 5:15am UTC](https://discuss.huggingface.co/t/how-should-i-balance-learning-rate-and-data-sampling-during-cpt-on-multiple-datasets/177517 "2026-07-07T05:15:59Z")
Hi everyone, I’m doing continued pretraining on a mixture of eight open-source datasets. The training is now in the late stage where the validation loss decreases very slowly. My objective is to minimize a weighted val… --- ## [\[Research\] From Functional Geometry to Dynamic Grammar: New LIMEN Audits (V23–V24) Across 7 Architectures](https://discuss.huggingface.co/t/research-from-functional-geometry-to-dynamic-grammar-new-limen-audits-v23-v24-across-7-architectures/177260)
**Author:** [@jeanbatuli](https://discuss.huggingface.co/u/jeanbatuli)\ **Replies:** 2\ **Last updated:** [July 2, 2026, 3:17pm UTC](https://discuss.huggingface.co/t/research-from-functional-geometry-to-dynamic-grammar-new-limen-audits-v23-v24-across-7-architectures/177260 "2026-07-02T15:17:22Z")
Hi everyone, I am sharing recent results from my independent research project, LIMEN (Liminal Internal Metric for Emergent Navigation), which aims to characterize the internal dynamics of Transformers through hidden sta… --- ## [A comprehensive, bilingual guide to Transformers: From foundations to KV-cache compression & attention dynamics](https://discuss.huggingface.co/t/a-comprehensive-bilingual-guide-to-transformers-from-foundations-to-kv-cache-compression-attention-dynamics/177222)
**Author:** [@karlexmarin](https://discuss.huggingface.co/u/karlexmarin)\ **Replies:** 0\ **Last updated:** [June 29, 2026, 3:54am UTC](https://discuss.huggingface.co/t/a-comprehensive-bilingual-guide-to-transformers-from-foundations-to-kv-cache-compression-attention-dynamics/177222 "2026-06-29T03:54:17Z")
Hi everyone, I’d like to share an open-source resource I’ve been working on: a comprehensive, bilingual (English & Spanish) guide on Transformer architectures. My goal was to create a bridge between the mathematical fo… --- ## [Error fix of the 503 loop](https://discuss.huggingface.co/t/error-fix-of-the-503-loop/177164)
**Author:** [@ronnie-katende](https://discuss.huggingface.co/u/ronnie-katende)\ **Replies:** 1\ **Last updated:** [June 25, 2026, 10:18am UTC](https://discuss.huggingface.co/t/error-fix-of-the-503-loop/177164 "2026-06-25T10:18:42Z")
Title: Space stuck in 503 loop — restart and commits not working Body: My Space is stuck in a 503 error and cannot be restarted. I have tried: Clicking “Restart Space” and “Factory Rebuild” from the UI Pushing multi… [Next page](https://discuss.huggingface.co/c/transformers/9.md?page=1)