selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

fastino logo

Can the Tesla T4 run GLiDE-27B?

GLiDE-27B·26.1B·Tesla T4·16 GB VRAM

Saved in this browser, used on every model page.
Yes, at low quality

Yes, but only at Q3_K_M (about 15 GB of 16 GB), which costs noticeable quality, at roughly 14 tokens/s.

best quant
Q3_K_M
needs
15 GB
usable
16 GB
est. speed
~14 tok/s

Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the GLiDE-27B page →

Every quantisation of GLiDE-27B on the Tesla T4

QuantQualityMemoryFitSpeed
Q2_K Heavy loss11 GBfits~21 tok/s
Q3_K_M pickNoticeable loss15 GBfits~14 tok/s
Q4_K_M Small loss19 GBoffload~6.9 tok/s
Q5_K_M Near-lossless21 GBoffload~4.8 tok/s
Q6_K Near-lossless24 GBoffload~3.4 tok/s
Q8_0 Lossless30 GBoffload~2.1 tok/s
FP16 / BF16 Lossless54 GBtoo big—

Speed is a bandwidth-based estimate for generating text (320 GB/s); prompt processing and long contexts are slower.

Measured speed on the Tesla T4

No community reports for this pair yet.

Report your own speed →

Run GLiDE-27B on the Tesla T4

Using Q3_K_M, the best fit for Tesla T4.

llama.cpp · OpenAI-compatible server with a web UIdocs
llama-server -hf kingjones777/GLiDE-27B-ROCmFP4-STRIX-GGUF:Q3_K_M

Downloads the GGUF on first run. Installed through llama.app? The same flags work with `llama serve`.

vLLM · Serve an OpenAI-compatible APIdocs
vllm serve fastino/GLiDE-27B
vLLM · Docker Compose (NVIDIA)docs
services:
  vllm:
    image: vllm/vllm-openai:latest
    command: fastino/GLiDE-27B
    ports:
      - "8000:8000"
    ipc: host
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

API on port 8000.

SGLang · Serve an OpenAI-compatible APIdocs
sglang serve fastino/GLiDE-27B --host 0.0.0.0 --port 30000
SGLang · Docker Compose (NVIDIA)docs
services:
  sglang:
    image: lmsysorg/sglang:latest
    command: sglang serve fastino/GLiDE-27B --host 0.0.0.0 --port 30000
    ports:
      - "30000:30000"
    ipc: host
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

API on port 30000.

Transformers · Ask about an imagedocs
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="fastino/GLiDE-27B", dtype="auto", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
            {"type": "text", "text": "Describe this image."},
        ],
    }
]
out = pipe(text=messages, max_new_tokens=256)
print(out[0]["generated_text"][-1]["content"])

Vision models that fit the Tesla T4 comfortably

All that fit →
ModelParamsBest quantMemorySpeedLicence
12.2BQ6_K≈13 GBFits in memory~18 tok/sConditions
19.4BA3BQ4_K_M≈13 GBFits in memory~97 tok/sCommercial OK
9.7BQ8_0≈11 GBFits in memory~17 tok/sCommercial OK
13.2BQ6_K≈13 GBFits in memory~16 tok/sConditions
8.8BQ8_0≈10 GBFits in memory~19 tok/sCommercial OK
10.7BQ8_0≈12 GBFits in memory~16 tok/sConditions

Other GPUs that run GLiDE-27B

Frequently asked questions

Can the Tesla T4 run GLiDE-27B?
Yes, but only at Q3_K_M (about 15 GB of 16 GB), which costs noticeable quality, at roughly 14 tokens/s.
How fast does GLiDE-27B run on the Tesla T4?
An estimate from memory bandwidth gives about 14 tokens/s at Q3_K_M.
Which quantisation of GLiDE-27B should I use on the Tesla T4?
Q3_K_M: the largest one that fits. Expected quality: noticeable loss compared with the full-precision weights.

Memory and speed are estimates from the model's size and the card's memory bandwidth; real numbers depend on the runtime, drivers and context length. Community figures are medians of user reports.