Problemsby GlyphDeepSeek users keep filing bugs
18 posts from 18 people15d activeSteady, +3 this week
Posts per day
The posts behind it
[Feature]: Integrate FlashTopK fused indexer top-k for DSA prefill on Hopper (SM90)
### 🚀 The feature, motivation and pitch [FlashTopK](https://github.com/Tencent-Hunyuan/FlashTopK) (Apache-2.0) is an exact, deterministic fused indexer-scoring + top-k kernel for DSA-style sparse attention on Hopper (SM90), written mostly in CuTe DSL. Its README reports...
Continuous batching: model support tracker
Tracking which model families continuous batching (`generate_batch`, `ContinuousBatchingManager`) runs. The engine takes its attention types from `layer_types` (`group_layers_by_attn_type`) and sizes its paged cache from `num_key_value_heads` and `head_dim`; anything outside...
ollama-cloud: deepseek-v4-pro rejects max_tokens > 65536 despite model's ~384k output capability
**Note: this is an Ollama Cloud API issue, not a local runtime issue - the OS/GPU/CPU fields below are intentionally left empty.** ## What is the issue? Ollama Cloud (`https://ollama.com/v1`, OpenAI-compatible API) rejects chat completion requests to `deepseek-v4-pro` and...
[RFC]: Let replicas read each other's GPU KV cache (L1 sharing)
> [!NOTE] > This RFC is at an early stage and under active development. We will keep updating the design and results as the work progresses, and feedback is very welcome. Terms: **L1** = a replica's GPU KV cache; **L2** = host-DRAM KV tier; **L3** = SSD or remote KV store. ###...
[Feature]: Add an NVIDIA RTX PRO 6000 Blackwell deployment recipe for DeepSeek-V4.1-Flash
### 🚀 The feature, motivation and pitch Target repo: vllm-project/recipes (content behind https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash) Related model page: https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4.1-Flash CPU：AMD Epyc 9355 x2 GPU：Nvidia RTX PRO 6000 8GPU...
[Bug]: DeepSeek-OCR/DeepSeek-OCR-2 crashes on a zero-sum tiled image
### Your current environment - vLLM: `0.1.dev23679+gaaa67b754`, editable checkout at `aaa67b754ebe26247d0b97c8f412397eac066184` (clean working tree; precompiled wheel artifacts) - OS: Ubuntu 26.04.1 LTS, Linux `7.0.0-38-generic`, x86_64, glibc 2.43 - GPU: NVIDIA GeForce RTX 5070...
[Bug][KV Offload]: Direct-layout offload is treated as parallelism-agnostic, so engines with different TP sizes sharing a secondary tier load corrupted KV
### Your current environment 4×H200, vLLM nightly `0.32.0rc1.dev100`, model `meta-llama/Llama-3.2-1B`, `OffloadingConnector` with `TieringOffloadingSpec` and a `p2p` or `fs` secondary tier. ### 🐛 Describe the bug When two engines with **different TP sizes** share a secondary...
[RFC / Research Showcase] SPICY-KV: 28.6x–50x Pre-RoPE Algebraic Multigrid KV Compression + 4-Byte Subgrid Variance Calibration (91.7%–100% RULER Recall on DeepSeek-R1 7B & 1.5B)
### Motivation. ### Summary & Motivation for Long-CoT / PagedAttention Serving Hi `vLLM` team! Sharing our newly released paper, `RULER` benchmarks, and reference implementation for **SPICY-KV** (*SPectral Invariant Coarse-graining with dYadic KV renormalization*), which...
the websearch is disabled in random models
### What is the issue? deepseek 4.1 flash-cloud, some qwen3.5 models, cannot pop up websearch tool. Restart the app or mac, does not solve the problem. ### Relevant log output ### OS _No response_ ### GPU _No response_ ### CPU _No response_ ### Ollama version _No response_
Grok Bot worth it when I already run an autonomous agent (Hermes + DeepSeek) on my own VPS with zero bottleneck?
...time. 30+ scheduled tasks total.. The honest situation: I have NOT hit any performance bottleneck. CPU and RAM. are fine, tasks finish when scheduled, and per-task cost is tiny because. DeepSeek is basically free compared to hosted agents.. What I've gathered about Grok Bot so...
How is it possible to enable tool calling like file read/write/execute in Strata (Qwen3.8-next-flash) web ui?
...Strada and Qwen3.8-next-flash on my PC and I was amazed by it.. The problem is, I cant seem to find a way to enable or add tool callings, so it can read/write/execute files on my PC. How do you do that in Strata?. Thank you very much in advance. Update:. Per suggestions, I went...
[Bug]: GLM-5.3-Flash JIT-compiles DeepGEMM `tf32_hc_prenorm_gemm` during serving (20 mHC split-K variants are never warmed)
### Your current environment The output of python collect_env.py - vLLM v0.31.0 (official image, banner `version 0.31.0`) - 1x NVIDIA B300 SXM6 AC (SM103, 148 SMs), TP=1 - GLM-5.3-Flash, NVFP4 checkpoint (`nvidia/GLM-5.3-Flash-NVFP4`), `hidden_size=4096`, `hc_mult=4` ### 🐛...
[Bug]: [Bug]: deepseek_r1 streaming reasoning deltas are not detokenized (byte-level BPE glyphs)
### Your current environment The output of python collect_env.py ### 🐛 Describe the bug ### Describe the bug With `--reasoning-parser deepseek_r1` and `"stream": true`, the reasoning deltas are emitted as raw byte-level BPE token glyphs instead of decoded UTF-8. `delta.content`...
[Bug]: DeepSeek-V4.1-Flash cannot run on SM120 (RTX PRO 6000): compressed-layer page_block_size=32 has no FlashInfer sparse-MLA decode kernel instantiation (only pbs=64)
...`--tokenizer-mode deepseek_v41` ### 🐛 Describe the bug DeepSeek-V4.1-Flash fails to start on SM120 with: ### Root cause analysis V4.1's two-tier sparse attention gives each layer a `compress_ratio` (this checkpoint: `[0,0,2,2,...,2,1,1,1,1,1]`). vLLM computes the...
[Bug]: Derender drops the leading space on SentencePiece tokenizers
### Your current environment Tokenizer only, so it's not platform specific. Reproduced on `main` (a2be4d3cc7) with `tokenizers` 0.22. ### 🐛 Describe the bug On SentencePiece/Metaspace tokenizers, `/v1/chat/completions/derender` and `/v1/completions/derender` drop the first...
[langchain-deepseek] bind_tools(tool_choice="any") produces an invalid request with DeepSeek's default thinking mode`
### Submission checklist - [x] This is a bug, not a usage question. - [x] I added a clear and descriptive title that summarizes this issue. - [x] I used the GitHub search to find a similar question and didn't find it. - [x] I am sure that this is a bug in LangChain rather than...
[Model Support] DeepSeek-V4.1 Tracking Issue
Model: [deepseek-ai/DeepSeek-V4.1-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) Recipe: [`models/deepseek-ai/DeepSeek-V4.1-Flash.yaml`](https://github.com/vllm-project/recipes/blob/main/models/deepseek-ai/DeepSeek-V4.1-Flash.yaml) - verified on H200, GB200,...
[Feature]: DeepSeek-V4.1-Flash Kernels Integration & Optimization Tracker
### 🚀 The feature, motivation and pitch DeepSeek releases some kernels for [deepseek-ai/DeepSeek-V4.1-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) We are currently integrating these kernels into vLLM. -...
Companies and products named
About this problem
Evidence
18 posts from 18 people in 6 places, about 2 a week over 58 days. Mostly on vllm-project/vllm, ollama/ollama, langchain-ai/langchain. Tools named alongside: vLLM, NVIDIA, GitHub, Hugging Face.
Frustration Frustration 1 of 3· Seen on GitHub issues, Reddit
How it was grouped
Posts about DeepSeek that state a pain but match no known issue yet. First post Sep 10, 2026, latest Nov 6, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
Other DeepSeek problems
Running DeepSeek locally hits memory and speed limits
8 posts from 8 people on GitHub issues, Hacker News.
DeepSeek GPU access and drivers are painful
7 posts from 6 people on GitHub issues.
Trouble with DeepSeek cloud
Top post: "Request: Add cloud support for DeepSeek-V4.1-Flash", one of 5 posts from 5 people on GitHub issues.
DeepSeek rate limits and usage caps block work
3 posts from 3 people on Hacker News, GitHub issues, Reddit.
Trouble with DeepSeek flash
Top post: "Asymmetric P/D Deployment for DeepSeek-V4.1 Flash", one of 3 posts from 3 people on GitHub issues, Community forums.
DeepSeek lock-in makes switching costly
2 posts from 2 people on GitHub issues. Unverified.
DeepSeek upgrades break working setups
2 posts from 2 people on GitHub issues. Unverified.
DeepSeek fails to return valid structured output
2 posts from 2 people on GitHub issues. Unverified.
History