Skip to content

[Issue]: degraded qwen3.6-moe modell quality on FLM v1.0.2 #679

Description

@benoe77

Problem Description

Environment & Context

  • Working Baseline: FastFlowLM v1.0.1 (serving by lemonade) using previously downloaded model weights
  • Regressed Setup: FastFlowLM v1.0.2 (.deb) freshly installed, using newly downloaded model weights/quantization (qwen3.6-moe:35b-a3b).

Summary

When benchmarking the newly updated model weights under FastFlowLM v1.0.2 against the earlier v1.0.1 setup, the new environment exhibits severe reasoning instability, infinite thinking loops, language drift (switching to German on Hungarian prompts), and token degeneration (///) under complex/constrained prompts.


Key Comparison

Metric Previous (FLM v1.0.1 + Old Weights) New (FLM v1.0.2 + New Weights)
Total Benchmark Time ~27 minutes (1,649s) ~86 minutes (5,164s)
Total Tokens Generated 22,521 tokens 75,706 tokens
Generation Speed ~13.7 t/s (Consistent) ~14.2 – 17.6 t/s (Volatile)
Loop / Degeneration Rate 0 / 5 (All completed cleanly) 2 / 5 (Hit max token limits)
Language Adherence 100% (Hungarian) Failed (Drifted to German in Q3)

Detailed Symptoms

1. Infinite Reasoning Loops (<think> block failure)

On complex reasoning/math prompts, the model under v1.0.2 enters an endless loop inside the internal <think> block. It repeatedly recalculates and re-evaluates the same steps without reaching the </think> closing tag until it exhausts the maximum generation limit (32,640 tokens).

2. Output Degeneration under Constraints

On strict constraint-following tasks (e.g., creative writing without specific letters), the generation breaks down completely. After generating thousands of tokens in a loop, the output degrades into an endless repetition of forward slashes (///) until hitting the token cap.

3. Language Drift / Loss of System Context

On code optimization tasks, the model completely loses target language context during the <think> reasoning phase. Despite the user prompt being entirely in Hungarian, the final generated explanation and section headers unexpectedly output in German (e.g., "Python Implementierung", "Algorithmus-Erklärung").

4. Wasted Compute and Inflated Runtime

Even though the raw generation speed of the new engine was slightly higher (~17.5 t/s vs ~13.7 t/s), the total benchmark runtime exploded by +213% (~86 mins vs ~27 mins) due to tens of thousands of wasted tokens generated inside reasoning loops.


Specific Case Comparison

  • Math/Probability Prompt:

    • v1.0.1: Completed successfully in 332s (4,552 tokens).
    • v1.0.2: Got stuck in a reasoning loop, ran for 2,290s, and hit the 32,640 token limit.
  • Code Optimization Prompt:

    • v1.0.1: Completed in 251s in Hungarian as requested.
    • v1.0.2: Ran for 295s, but unexpectedly switched to German explanations.
  • Constraint Following Prompt:

    • v1.0.1: Completed in 598s (8,106 tokens), correctly adhering to rules.
    • v1.0.2: Ran for 2,279s, hit the 32,640 token limit, and degraded into continuous /// repetition.

Expected Behavior

The model/engine under v1.0.2 should:

  1. Correctly terminate reasoning blocks (</think>) without looping infinitely.
  2. Maintain system context and strictly adhere to the prompt's language (no cross-lingual drift).
  3. Respect strict decoding constraints without falling into character degeneration (///).

Operating System

Ubuntu 26.04 LTS (Resolute Raccoon)

CPU

AMD Ryzen AI 7 350 w/ Radeon 860M

GPU

AMD Ryzen AI 7 350 w/ Radeon 860M

ROCm Version

7.14.0

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

rocminfo --support output
Paste output here

Additional Information

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions