[Bug] GLM-5.2 with MTP+Hicache+Mooncake occasionally produces garbled output on latest SGLang

### Checklist - [x] I searched related issues but found no solution. - [x] The bug persists in the latest version. - [x] Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback. - [x] If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed. - [ ] Please use English. Otherwise, it will be closed. ### Describe the bug ## Bug: GLM-5.2 produces garbled / corrupted text in long-tail generation cases ### Summary When serving GLM-5.2 with SGLang, I occasionally encounter garbled output during generation, especially in long-tail cases. The model starts normally, but after a certain point the generated text may contain corrupted characters, repeated abnormal tokens, or unreadable mixed text. I am not sure whether this is caused by tokenizer handling, chat template formatting, sampling settings, model compatibility, or a known issue in the current SGLang runtime. I would like to ask whether the community has seen similar behavior and whether there is a recommended workaround. ### Environment - SGLang version: `latest / main branch` - GPU: `B300` - Model: GLM-5.2 ### Reproduction The issue is intermittent and appears more often on long-tail prompts or longer generations. <img width="786" height="494" alt="Image" src="https://github.com/user-attachments/assets/43981185-105d-4d97-bc68-07ce99e707d7" /> <img width="1631" height="513" alt="Image" src="https://github.com/user-attachments/assets/48abdfb8-1911-448c-ab5b-2f124dcd4543" /> ### Reproduction The issue is intermittent, but it is easier to reproduce with long-tail prompts or long generations when serving GLM-5.2-FP8 on the latest SGLang with MTP/EAGLE speculative decoding and Mooncake hierarchical cache enabled. My deployment setup: - 4 machines - 8 GPUs per machine - 2 SGLang servers per machine: - `CUDA_VISIBLE_DEVICES=0,1,2,3`, port `30000` - `CUDA_VISIBLE_DEVICES=4,5,6,7`, port `30001` - Each SGLang server uses `tp-size=4` - Model path: `/modules/GLM-5.2-FP8` - Speculative decoding / MTP enabled: - `--speculative-algorithm EAGLE` - `--speculative-num-steps 5` - `--speculative-eagle-topk 1` - `--speculative-num-draft-tokens 6` - Mooncake hierarchical cache enabled: - `--enable-hierarchical-cache` - `--hicache-storage-backend mooncake` - `--hicache-write-policy write_through` - `--hicache-storage-prefetch-policy timeout` - `--hicache-mem-layout page_first_direct` - KV cache dtype: `fp8_e4m3` Example launch command for one SGLang instance: ```bash CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m sglang.launch_server \ --model-path /modules/GLM-5.2-FP8 \ --host 0.0.0.0 \ --port 30000 \ --trust-remote-code \ --dist-init-addr 127.0.0.1:5501 \ --nnodes 1 \ --tp-size 4 \ --speculative-algorithm EAGLE \ --speculative-num-steps 5 \ --speculative-eagle-topk 1 \ --speculative-num-draft-tokens 6 \ --served-model-name glm5.2-jz \ --tool-call-parser glm47 \ --reasoning-parser glm45 \ --mem-fraction-static 0.90 \ --kv-cache-dtype fp8_e4m3 \ --chunked-prefill-size 32768 \ --max-prefill-tokens 32768 \ --cuda-graph-max-bs 16 \ --max-running-requests 100 \ --enable-request-time-stats-logging \ --enable-metrics \ --enable-cache-report \ --schedule-conservativeness 0.1 \ --stream-interval 1 \ --stream-response-default-include-usage \ --model-loader-extra-config '{"enable_multithread_load": true}' \ --enable-hierarchical-cache \ --hicache-ratio 2 \ --hicache-write-policy write_through \ --hicache-storage-prefetch-policy timeout \ --hicache-mem-layout page_first_direct \ --hicache-storage-backend mooncake \ --hicache-storage-backend-extra-config '{"standalone_storage": false, "master_server_address": "10.218.104.1:53011", "metadata_server": "http://10.218.104.1:8011/metadata", "local_hostname": "10.218.104.1", "global_segment_size": "0", "protocol": "rdma", "device_name": "mlx5_0,mlx5_4,mlx5_6,mlx5_14", "check_server": true, "master_metrics_port": 9011}' Mooncake master/client are also enabled with RDMA and offload: mooncake_master \ --rpc_port=53011 \ --enable_http_metadata_server=true \ --http_metadata_server_port=8011 \ --metrics_port=9011 \ --eviction_high_watermark_ratio=0.95 \ --allocation_strategy=free_ratio_first \ --enable_offload=true mooncake_client \ --host=<machine_ip> \ --port=52011 \ --global_segment_size=250GB \ --master_server_address=10.218.104.1:53011 \ --metadata_server=http://10.218.104.1:8011/metadata \ --protocol=rdma \ --device_names=<mlx5 devices> \ --enable_offload=true After the servers are up, I send long-tail prompts or long-generation requests through the OpenAI-compatible API. The response often starts normally, but in some requests the later part becomes garbled/corrupted, with unreadable characters, abnormal repeated tokens, or semantically broken text. The issue is not 100% reproducible, but it appears more frequently with long-tail prompts or longer generations under this MTP + Mooncake configuration. ``` ### Environment - SGLang version: latest / main branch - Model: GLM-5.2-FP8 - Model path: `/modules/GLM-5.2-FP8` - Serving API: OpenAI-compatible API