Skip to content

Instantly share code, notes, and snippets.

View fank's full-sized avatar

Florian Kinder fank

  • @mcl-de enthus GmbH
  • Germany
  • 11:10 (UTC +02:00)
View GitHub Profile
@fank
fank / repro_48645.py
Created July 14, 2026 19:37
Repro for vllm-project/vllm#48645 — deepseek_v4 parser routes answer to reasoning_content when the model omits </think>
#!/usr/bin/env python3
"""
Repro for vllm-project/vllm#48645 — deepseek_v4 parser: a reply that omits `</think>`
routes the entire answer to reasoning_content, leaving content empty (trailing EOS un-stripped).
Server must run (as the DeepSeek-V4-Flash recipe suggests):
--reasoning-parser deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice
--default-chat-template-kwargs={"thinking":true,"reasoning_effort":"high"}
The request deliberately sends NO chat_template_kwargslike any ordinary OpenAI-compatible
@fank
fank / MI50-ALL-TESTS-master-log-2026-07-03.md
Created July 4, 2026 14:44
MI50 (gfx906) LLM-Serving

Card: AMD Instinct MI50 32 GB (gfx906 / Vega 20 / GCN5, ~1 TB/s HBM2, no matrix cores, passive-cooled) · replaced an Intel Arc B580 (permanent swap) Host: Nobara/Fedora, Ryzen 7 8845HS (8c/16t), 30 GB RAM · SELinux disabled · Docker Workload: n8n LLM node — "big data in → small data out" (prefill-dominated), single-user (--parallel 1) Compiled: 2026-07-03 — consolidates every test from 2026-06-20 → 2026-07-03. Source reports: MI50-report.md, MI50-MTP-speculative-decoding-analysis.md, MI50-Q5-Q6-quant-test.md, DFlash-setup.md, dflash-bench-2026-06-29.md, ornith-mtp-optimization-report.md, ornith-model-docker-logs-2026-07-01.md, llama-stats-2026-06-23.md, memory mi50-*.

Current production (2026-07-03): llama-hipgraphs container, image llama-hipgraphs:upstream-rocm-7.2.4 (build commit 0eca4d4, 2026-06-30), model Ornith-1.0-35B Q4_K_M + embedded MTP, --spec-type draft-mtp --spec-draft-n-max 2, 262 k ctx, q8_0 KV, --parallel 1, port 8089. ~70 t/s.

@fank
fank / klipper.conf
Created September 2, 2021 21:00
CR-5 Pro Klipper config
# This file contains common pin mappings for the BIGTREETECH SKR mini
# E3 v2.0. To use this config, the firmware should be compiled for the
# STM32F103 with a "28KiB bootloader". Also, select "Enable extra
# low-level configuration options" and configure "GPIO pins to set at
# micro-controller startup" to "!PA14".
# The "make flash" command does not work on the SKR mini E3. Instead,
# after running "make", copy the generated "out/klipper.bin" file to a
# file named "firmware.bin" on an SD card and then restart the SKR
# mini E3 with that SD card.
@fank
fank / sluice.cs
Last active January 13, 2017 18:20
Space Engineers Sluice Management with 2 doors, 2 lights, 1 Airvent and 1 Timer
public Program() {
// The constructor, called only once every session and
// always before any other method is called. Use it to
// initialize your script.
//
// The constructor is optional and can be removed if not
// needed.
}
public void Save() {
@fank
fank / beguidcalc.md
Last active August 29, 2024 17:52
ArmA 3 / DayZ-Standalone - BattlEye GUID calculation

md5("BE" (2 bytes) + 64-bit SteamID (8 bytes))