refactor : use common download in tools/run

[angt]

• update the common/download interface to be directly usable by tools/run (removing duplicated code).
• fix ollama downloads by implementing manual redirect handling (addressing issues with cpp-httplib).

[angt]

With this change, we can deprecate the cURL dependency and start shipping releases without it.
The unified HTTP stack will become a reality :)

[ericcurtin]

My reviews are somewhat irrelevant now, I don't have merge rights, took a quick skim didn't read every line, at a glance everything seems reasonably ok, recommend doing a quick test via:

llama-server -dr gemma3

to be sure...

[angt]

My reviews are somewhat irrelevant now, I don't have merge rights, took a quick skim didn't read every line, at a glance everything seems reasonably ok, recommend doing a quick test via:

llama-server -dr gemma3

to be sure...

Here some runs:

via docker

$ ./build/bin/llama-server -dr gemma3
ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices
ggml_metal_library_init: using embedded metal library
ggml_metal_library_init: loaded in 0.021 sec
ggml_metal_device_init: GPU name:   Apple M3
ggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)
ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)
ggml_metal_device_init: simdgroup reduction   = true
ggml_metal_device_init: simdgroup matrix mul. = true
ggml_metal_device_init: has unified memory    = true
ggml_metal_device_init: has bfloat            = true
ggml_metal_device_init: has tensor            = false
ggml_metal_device_init: use residency sets    = true
ggml_metal_device_init: use shared buffers    = true
ggml_metal_device_init: recommendedMaxWorkingSetSize  = 19069.67 MB
common_docker_resolve_model: Downloading Docker Model: ai/gemma3:latest
common_download_file_single_online: no previous model file found /Users/angt/Library/Caches/llama.cpp/ai_gemma3_latest.gguf
common_download_file_single_online: trying to download model from https://production.cloudflare.docker.com/registry-v2/docker/registry/v2/blobs/sha256/04/04a43a22e8d2003deda5acc262f68ec1005fa76c735a9962a8c77042a74a7d19/data?expires=1764234374&signature=m2%2BBuw6sCMTEH4cNizZDs6fsLC8%3D&version=2 to /Users/angt/Library/Caches/llama.cpp/ai_gemma3_latest.gguf.downloadInProgress (etag:"562fa5cb63ae8b96836b09a658443c01-25")...
^C==============================>                   ]  63%  (1503 MB / 2374 MB)

(resume works)

via HF:

$ ./build/bin/llama-server -hf unsloth/gpt-oss-120b-GGUF
ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices
ggml_metal_library_init: using embedded metal library
ggml_metal_library_init: loaded in 0.006 sec
ggml_metal_device_init: GPU name:   Apple M3
ggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)
ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)
ggml_metal_device_init: simdgroup reduction   = true
ggml_metal_device_init: simdgroup matrix mul. = true
ggml_metal_device_init: has unified memory    = true
ggml_metal_device_init: has bfloat            = true
ggml_metal_device_init: has tensor            = false
ggml_metal_device_init: use residency sets    = true
ggml_metal_device_init: use shared buffers    = true
ggml_metal_device_init: recommendedMaxWorkingSetSize  = 19069.67 MB
common_download_file_single_online: no previous model file found /Users/angt/Library/Caches/llama.cpp/unsloth_gpt-oss-120b-GGUF_gpt-oss-120b-F16.gguf
common_download_file_single_online: trying to download model from https://cas-bridge.xethub.hf.co/xet-bridge-us/68923b51e5822b89fab7a1e7/f7c8b3cdb2bacb3cef00372f4fa3070f250a2d8838cec29653b9b9db8238a583?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=cas%2F20251127%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20251127T081900Z&X-Amz-Expires=3600&X-Amz-Signature=555418210eb5445a4ca2376a0a0b4c15c8b5cef28cc67e4894397e8c41243075&X-Amz-SignedHeaders=host&X-Xet-Cas-Uid=public&response-content-disposition=inline%3B+filename*%3DUTF-8%27%27gpt-oss-120b-F16.gguf%3B+filename%3D%22gpt-oss-120b-F16.gguf%22%3B&x-id=GetObject&Expires=1764235140&Policy=eyJTdGF0ZW1lbnQiOlt7IkNvbmRpdGlvbiI6eyJEYXRlTGVzc1RoYW4iOnsiQVdTOkVwb2NoVGltZSI6MTc2NDIzNTE0MH19LCJSZXNvdXJjZSI6Imh0dHBzOi8vY2FzLWJyaWRnZS54ZXRodWIuaGYuY28veGV0LWJyaWRnZS11cy82ODkyM2I1MWU1ODIyYjg5ZmFiN2ExZTcvZjdjOGIzY2RiMmJhY2IzY2VmMDAzNzJmNGZhMzA3MGYyNTBhMmQ4ODM4Y2VjMjk2NTNiOWI5ZGI4MjM4YTU4MyoifV19&Signature=v02cqmRkLDiiyjkhUGUmvIzg9woZZbAsLWo9RCrqcYg6FB-wZboc4gLBqSvMjHzz8lCEVskrLy3ZrPPV8j%7E%7Ep%7EqqB04dzNIy608mlJUIUXm51Ux9%7EHYqjo9oZGi0ZJgoxNvnH5TU5Nn%7ELyftUrpB53b6BobQyzS66myRjsnEmIkYOyzXbZUBKRfpK9Hu65PUFvfAYEq0rC%7E6x4y8CLHu0eH8oX41UiKykrtZGiRGluYXztE0sBJpWGJERgp2Wv5L1qSX1J2YpgM7CJre0QPsk8xa8ZbP-lgRtvSj%7ELPkR3pwys4NxYillvJdB%7EI1uF-vURScMFi3COObp3OAqn-T6Q__&Key-Pair-Id=K2L8F4GPSG1IFC to /Users/angt/Library/Caches/llama.cpp/unsloth_gpt-oss-120b-GGUF_gpt-oss-120b-F16.gguf.downloadInProgress (etag:"f7c8b3cdb2bacb3cef00372f4fa3070f250a2d8838cec29653b9b9db8238a583")...
^C=========>                                        ]  21%  (13202 MB / 62340 MB)

and for tools/run:

$ ./build/bin/llama-run llama3
common_docker_resolve_model: Downloading Docker Model: library/llama3:latest
common_download_file_single_online: no previous model file found /Users/angt/Library/Caches/llama.cpp/library_llama3_latest.gguf
common_download_file_single_online: 403 on HEAD, assuming GET/Resume is allowed
common_download_file_single_online: trying to download model from https://dd20bb891979d25aebc8bec07b2b3bbc.r2.cloudflarestorage.com/ollama/docker/registry/v2/blobs/sha256/6a/6a0746a1ec1aef3e7ec53868f220ff6e389f6f8ef87a01d77c96807de94ca2aa/data?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=66040c77ac1b787c3af820529859349a%2F20251127%2Fauto%2Fs3%2Faws4_request&X-Amz-Date=20251127T082042Z&X-Amz-Expires=86400&X-Amz-SignedHeaders=host&X-Amz-Signature=891427f6f6e2b09cafa8869ca3d84993b1c17352b11f342e2b8c033578a77ca0 to /Users/angt/Library/Caches/llama.cpp/library_llama3_latest.gguf.downloadInProgress (etag:)...
^C=>                                                ]   4%  (185 MB / 4445 MB)

the output is excessively verbose, and the progress bar is broken when doing multiple downloads. However, it will be easier to improve the code from now

[CISC]

Looks like thread-safety test fails across the board.

[angt]

Looks like thread-safety test fails across the board.

Yes I’m going to check all the red alerts. 😬

[ericcurtin]

@angt I don't know if you are interested in this but, it would be preferable if llama-run ran a client and a llama-server in parallel, llama-run/llama-server would benefit from more re-usable code

[angt]

@angt I don't know if you are interested in this but, it would be preferable if llama-run ran a client and a llama-server in parallel, llama-run/llama-server would benefit from more re-usable code

I tried to preserve llama-run’s current behavior as my main goal was to move toward removing the cURL dependency and fix the tool when cURL is disabled. But I fully agree that rethinking llama-run could be useful :)

[ericcurtin]

@angt I don't know if you are interested in this but, it would be preferable if llama-run ran a client and a llama-server in parallel, llama-run/llama-server would benefit from more re-usable code

I tried to preserve llama-run’s current behavior as my main goal was to move toward removing the cURL dependency and fix the tool when cURL is disabled. But I fully agree that rethinking llama-run could be useful :)

A CVE in linenoise ended up scratching my itch, but I'd appreciate if you built this branch, gave the new llama-run experience a shot and provided feedback for future PRs:

#17554

[ericcurtin]

You might want to abandon run.cpp changes here (I don't know if the other parts should stay in this PR). The new PR is a much better experience and CVE free (with the removal of linenoise)

[angt]

@ericcurtin, this PR still resolves some issues with cpp-httplib and improves download.cpp. I believe it can be merged before the complete rewrite of llama-run to address current issues more efficiently.

[ericcurtin]

I've noticed a pattern where my PRs are often asked to wait for your changes to merge first. While I understand the need to avoid conflicts, constantly deferring my work can be discouraging. Could we try prioritizing based on readiness this time?

#16196 (comment)

I have no further changes to make to my pull request at this time:

#17554

In this case there isn't any significant conflict, the above PR makes no changes to download code.

[angt]

I apologize if you felt that way, but I must point out that the PR you mentioned happened before yours. 😅

Anyway, I was only suggesting we merge this one with yours, but more importantly, that they don’t solve the same problem.