experimentalmachines/Qwen2.5-Math-1.5B-Instruct-ExecuTorch

Qwen2.5-Math-1.5B-Instruct for ExecuTorch

ExecuTorch exports of Qwen/Qwen2.5-Math-1.5B-Instruct (revision aafeb0fc6f22) for on-device inference with the openweights Android app or any ExecuTorch 1.4.0 runtime.

Files

Windows in this repo: XNNPACK (CPU) at 2k to 32k; Vulkan (GPU) at 2k to 32k. The window is fixed inside the file: the runtime allocates the whole KV cache at load, so pick the largest window the device can hold (fits_phone_budget in each folder's config.json is the estimate against a 5 GB budget).

Backend	Target	File	Window	Size	Smoke test
XNNPACK (CPU)	any arm64	xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-2k.pte	2,048 tokens	1.11 GB	passed ("Paris")
XNNPACK (CPU)	any arm64	xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-4k.pte	4,096 tokens	1.11 GB	passed ("Paris")
XNNPACK (CPU)	any arm64	xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-8k.pte	8,192 tokens	1.12 GB	passed ("Paris")
XNNPACK (CPU)	any arm64	xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-16k.pte	16,384 tokens	1.14 GB	passed ("Paris")
XNNPACK (CPU)	any arm64	xnnpack/Qwen2.5-Math-1.5B-Instruct-8da4w-gptq-32k.pte	32,768 tokens	1.17 GB	passed ("Paris")
Vulkan (GPU)	any arm64	vulkan/Qwen2.5-Math-1.5B-Instruct-vulkan-8da4w-2k.pte	2,048 tokens	1.40 GB	structure checked (no host NPU runtime)
Vulkan (GPU)	any arm64	vulkan/Qwen2.5-Math-1.5B-Instruct-vulkan-8da4w-4k.pte	4,096 tokens	1.41 GB	structure checked (no host NPU runtime)
Vulkan (GPU)	any arm64	vulkan/Qwen2.5-Math-1.5B-Instruct-vulkan-8da4w-8k.pte	8,192 tokens	1.42 GB	structure checked (no host NPU runtime)
Vulkan (GPU)	any arm64	vulkan/Qwen2.5-Math-1.5B-Instruct-vulkan-8da4w-16k.pte	16,384 tokens	1.44 GB	structure checked (no host NPU runtime)
Vulkan (GPU)	any arm64	vulkan/Qwen2.5-Math-1.5B-Instruct-vulkan-8da4w-32k.pte	32,768 tokens	1.49 GB	structure checked (no host NPU runtime)

Tokenizer: tokenizer.json, copied unchanged from the source repo. Each backend folder has a config.json listing every window as a variant with the metadata the .pte reports, and an export-report-<window>.json per file with the full export record.

Memory

• XNNPACK (CPU) at 2,048 tokens: the KV cache costs 57,344 bytes per token (fp32), 117,440,512 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 4,096 tokens: the KV cache costs 57,344 bytes per token (fp32), 234,881,024 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 8,192 tokens: the KV cache costs 57,344 bytes per token (fp32), 469,762,048 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 16,384 tokens: the KV cache costs 57,344 bytes per token (fp32), 939,524,096 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 32,768 tokens: the KV cache costs 57,344 bytes per token (fp32), 1,879,048,192 bytes for the whole window, allocated in full when the model loads.
• Vulkan (GPU) at 2,048 tokens: the KV cache costs 57,344 bytes per token (fp32), 117,440,512 bytes for the whole window, allocated in full when the model loads.
• Vulkan (GPU) at 4,096 tokens: the KV cache costs 57,344 bytes per token (fp32), 234,881,024 bytes for the whole window, allocated in full when the model loads.
• Vulkan (GPU) at 8,192 tokens: the KV cache costs 57,344 bytes per token (fp32), 469,762,048 bytes for the whole window, allocated in full when the model loads.
• Vulkan (GPU) at 16,384 tokens: the KV cache costs 57,344 bytes per token (fp32), 939,524,096 bytes for the whole window, allocated in full when the model loads.
• Vulkan (GPU) at 32,768 tokens: the KV cache costs 57,344 bytes per token (fp32), 1,879,048,192 bytes for the whole window, allocated in full when the model loads.

How it was made

• XNNPACK (CPU) any arm64: ExecuTorch 1.4.0 export_llm: 8-bit dynamic activations and 4-bit weights in groups of 32, int8 per-channel embeddings, XNNPACK with extended ops, prefill chunk 2048, fp32 KV cache. Built by run 1.
• Vulkan (GPU) any arm64: ExecuTorch 1.4.0 export_llm: 8-bit dynamic activations and 4-bit weights in groups of 32, int8 per-channel embeddings, the Vulkan delegate, prefill chunk 2048, fp32 KV cache. Built by run 1.

License

A quantized derivative of Qwen/Qwen2.5-Math-1.5B-Instruct, distributed under the same terms (apache-2.0). The upstream license files are included unchanged: LICENSE.