fix: enable TheRock ROCm installation on Windows for Radeon RX GPUs by anditherobot · Pull Request #2093 · lemonade-sdk/lemonade

anditherobot · 2026-06-04T04:02:33Z

fix: enable TheRock ROCm installation on Windows for Radeon RX GPUs (#2050)

Summary

Enables TheRock installation for Radeon RX GPUs when Lemonade detects the GPU by a family architecture code such as gfx120X.

In the reported Windows RX 9060 XT case, Lemonade correctly identifies the card as gfx120X. That code means the RX 9000 / RDNA4 GPU family. However, backend_versions.json did not list gfx120X, so Lemonade skipped the TheRock install step.

Change

Adds the Radeon RX family architecture codes to the TheRock install metadata:

GPU Series	Architecture Code	TheRock Package
RX 5000	`gfx101X`	`gfx101X-dgpu`
RX 6000	`gfx103X`	`gfx103X-all`
RX 7000	`gfx110X`	`gfx110X-all`
RX 9000	`gfx120X`	`gfx120X-all`

Also adds RX 5000 name detection so cards such as RX 5700 XT can resolve to gfx101X.

Testing

Confirmed locally on an RX 5700 XT that the ROCm backend is marked installable.
Verified the new TheRock package URLs exist on AMD's servers.
Confirmed the change covers the reported RX 9060 XT case, where detection returns gfx120X.

Fixes #2050

…k#2050) Allows Lemonade to automatically detect and download ROCm graphics acceleration files for AMD Radeon gaming graphics cards (RX 5000, 6000, 7000, and 9000 series) on Windows. Previously, these cards were recognized as unsupported, forcing the server to run text and image models on the slower CPU or Vulkan fallbacks.

superm1

I think this needs more matches for Linux too or this will cause an issue when a Linux user tries.

anditherobot · 2026-06-04T21:06:33Z

Good catch. I originally focused on Windows (from the original issue), but you're right that Linux needs this
too.
I mapped the specific Linux target codes (gfx1010/11/12) to gfx101X so the recipes and ROCm downloads
resolve them correctly. I also added unit tests for these Linux values.

LMK if this looks good.

superm1 · 2026-06-05T02:57:11Z

+    // RDNA1 family. Linux KFD reports the exact GPU ISA (gfx1010/gfx1011/gfx1012),
+    // while Lemonade recipes and ROCm backend downloads use the shared family name gfx101X.
+    {"gfx1010", "gfx101X"}, // Navi 10 (e.g., RX 5700 / 5700 XT)
+    {"gfx1011", "gfx101X"}, // Navi 12 (e.g., Radeon Pro 5600M)
+    {"gfx1012", "gfx101X"}, // Navi 14 (e.g., RX 5500 / 5500 XT)
+


you're adding these, but the mappings you added also had more (103x, 110x, 120x). Either add them all here or drop the mapping from backend_versions.json

superm1 · 2026-06-05T02:57:24Z

    }},
    {"llamacpp", "rocm", {"windows", "linux"}, {
-        {"amd_gpu", {"gfx1150", "gfx1151", "gfx103X", "gfx110X", "gfx120X"}},  // STX iGPUs + RDNA2/3/4 dGPUs
+        {"amd_gpu", {"gfx101X", "gfx1150", "gfx1151", "gfx103X", "gfx110X", "gfx120X"}},  // STX iGPUs + RDNA1/2/3/4 dGPUs


same comment as above, missing more mappings

Keep Linux ROCm detection consistent with the backend package mappings by normalizing gfx1033, gfx1035, and gfx1036 to gfx103X. Update the ROCm mapping test to cover the same behavior.

anditherobot · 2026-06-05T03:59:54Z

Thanks, this clarified the missing layer. Linux raw gfx... codes need translation before package selection.

I added gfx1033, gfx1035, and gfx1036 to the RDNA2 translations and updated the test.

Mapping for the codes covered by this PR:

GPU family	Linux raw code	Detection result	Recipe target	Download package
RDNA1	`gfx1010` / `gfx1011` / `gfx1012`	`gfx101X`	`gfx101X`	`gfx101X-dgpu`
RDNA2	`gfx1030` - `gfx1036`	`gfx103X`	`gfx103X`	`gfx103X-all`
RDNA3	`gfx1100` - `gfx1103`	`gfx110X`	`gfx110X`	`gfx110X-all`
RDNA3.5	`gfx1150` / `gfx1151` / `gfx1152`	exact code	exact code	exact package
RDNA4	`gfx1200` / `gfx1201`	`gfx120X`	`gfx120X`	`gfx120X-all`

Does this cover the layer mismatch you were pointing at, or am I still missing a case?

superm1 · 2026-06-05T04:12:17Z

All of the targets and how they map are described here: https://github.com/ROCm/TheRock/blob/main/cmake/therock_amdgpu_targets.cmake

You can cross reference this.

Completes RDNA3.5 mappings in backend versions and system detection, following ROCm's therock_amdgpu_targets.cmake target definitions.

anditherobot · 2026-06-05T05:14:26Z

Thanks for the reference! I cross-referenced therock_amdgpu_targets.cmake and added mapping support/unit tests for the missing gfx1153 (Radeon 820M) architecture target.

I hope that this fully covers the RDNA3.5 family. Here is the updated mapping table:

GPU family	Linux raw code	Detection result	Recipe target	Download package
RDNA1	`gfx1010` / `gfx1011` / `gfx1012`	`gfx101X`	`gfx101X`	`gfx101X-dgpu`
RDNA2	`gfx1030` - `gfx1036`	`gfx103X`	`gfx103X`	`gfx103X-all`
RDNA3	`gfx1100` - `gfx1103`	`gfx110X`	`gfx110X`	`gfx110X-all`
RDNA3.5	`gfx1150` / `gfx1151` / `gfx1152` / `gfx1153`	exact code	exact code	exact package
RDNA4	`gfx1200` / `gfx1201`	`gfx120X`	`gfx120X`	`gfx120X-all`

superm1 · 2026-06-05T21:53:58Z

    }},
    {"llamacpp", "rocm", {"windows", "linux"}, {
-        {"amd_gpu", {"gfx1150", "gfx1151", "gfx103X", "gfx110X", "gfx120X"}},  // STX iGPUs + RDNA2/3/4 dGPUs
+        {"amd_gpu", {"gfx101X", "gfx1150", "gfx1151", "gfx1152", "gfx1153", "gfx103X", "gfx110X", "gfx120X"}},  // STX iGPUs + RDNA1/2/3/4 dGPUs


You should just drop the comments. This is never going to age well. You also added things like Phoenix, hawk point here.

superm1 · 2026-06-07T14:58:38Z

 13. **Desktop app is on-demand; `lemond` runs independently** — On Windows, `LemonadeServer.exe` (which embeds `lemond` + tray icon) is the always-on process, auto-started via the Windows startup folder. The Tauri desktop app (`lemonade-app.exe`) is opened on demand when the user wants the UI and must not be added to startup. The desktop app must not embed or manage `lemond`'s lifecycle — it discovers the already-running server (UDP beacon for local, explicit base URL for remote) and speaks to it over HTTP.
+14. **AMD GPU platform parity** — When adding or expanding AMD/ROCm GPU support, keep Windows and Linux detection in sync. Windows commonly detects GPUs from product names like `5700` or `6800`, while Linux may report raw KFD ISA codes like `gfx1010` or `gfx1030`. Update `ROCM_ARCH_MAPPING` in `system_info.cpp` as needed so Linux detections resolve to the backend target expected by recipes and ROCm downloads, or to `""` when that ISA is intentionally unsupported.
+
+## Pending Local Work


Shouldn't this not commit?

superm1 · 2026-06-07T14:59:04Z

      "gfx908",
      "gfx90a",
      "gfx942",
+      "gfx101X",


I don't think this is correct here

superm1 · 2026-06-07T14:59:33Z


        std::string arch = lemon::SystemInfo::get_rocm_arch();
-        if (arch == "gfx1151") {
+        if (arch == "gfx1151" || arch == "gfx1152") {


Fair point. I don't have a way to test if gfx1152 actually needs this memory size fix.

yes please drop then.

anditherobot · 2026-06-07T21:17:29Z

Hi @superm1,

I did some homework From what I understand, gfx1153 (Radeon 820M) is listed in therock_amdgpu_targets.cmake, but its package currently returns 403 on AMD's download server. Mapping it to itself in our code will cause a download crash.

GPU	Linux KFD	Package	Status
`gfx1010` to `12`	`100100` to `100102`	`gfx101X-dgpu`	✅
`gfx1030` to `34`	`100300` to `100304`	`gfx103X-all`	✅
`gfx1033/35/36`	`100303` to `100306`	`gfx103X-all`	✅
`gfx1100` to `02`	`110000` to `110002`	`gfx110X-all`	✅
`gfx1103`	`110003`	`gfx110X-all`	✅
`gfx1150`	`110500`	`gfx1150`	✅
`gfx1151`	`110501`	`gfx1151`	✅
`gfx1152`	`110502`	`gfx1152`	✅
`gfx1153`	`110503`	`gfx1153`	❌ 403
`gfx1200/01`	`120000` to `120001`	`gfx120X-all`	✅

(Note: Generic gfx115X-all and gfx115X-igpu packages also return 403 — it seems AMD packages RDNA 3.5 integrated chips individually, not as a family).

I see two options to handle this:

A: Map gfx1153 → "" (empty value) — skip the ROCm graphics download and fall back to CPU mode.
(edits in system_info.cpp, backend_versions.json, and test_rocm_arch_mapping.py to remove gfx1153 mappings).
B: Map gfx1153 → gfx1152 — workaround the driver into loading the Kraken Point package.
(edit backend_versions.json to map the download package to gfx1152 and edit llamacpp_server.cpp to set HSA_OVERRIDE_GFX_VERSION=11.5.2 ).

What do you think of these approaches ?

(Edit: Replaced range dashes "–" with "to" in the table to clarify that these represent ranges of separate decimal target values, not a hyphenated integer format).

superm1 · 2026-06-08T00:15:06Z

Option A is the correct approach. We rely upon AMD binaries, and if AMD doesn't publish them we should fall back to vulkan or CPU.

superm1 · 2026-06-08T00:17:03Z

gfx1010–12 100100–02

HTH do you get 100100-02 from a 32 bit integer? This is impossible.

https://github.com/torvalds/linux/blob/master/drivers/gpu/drm/amd/amdkfd/kfd_topology.c#L472

anditherobot · 2026-06-08T00:46:34Z

gfx1010–12 100100–02

HTH do you get 100100-02 from a 32 bit integer? This is impossible.

https://github.com/torvalds/linux/blob/master/drivers/gpu/drm/amd/amdkfd/kfd_topology.c#L472

The dash was meant as range shorthand, not as a literal kernel value: gfx1010 / gfx1011 / gfx1012 =
100100 / 100101 / 100102

superm1 · 2026-06-08T01:04:40Z

gfx1010–12 100100–02

HTH do you get 100100-02 from a 32 bit integer? This is impossible.

https://github.com/torvalds/linux/blob/master/drivers/gpu/drm/amd/amdkfd/kfd_topology.c#L472

The dash was meant as range shorthand, not as a literal kernel value: gfx1010 / gfx1011 / gfx1012 =
100100 / 100101 / 100102

Ah Thanks for confirm

anditherobot mentioned this pull request Jun 4, 2026

therock installation does not work on windows #2050

Open

fl0rianr requested a review from superm1 June 4, 2026 12:03

superm1 reviewed Jun 4, 2026

View reviewed changes

Map Linux RDNA1 ROCm architectures

b096471

superm1 reviewed Jun 5, 2026

View reviewed changes

fix: map missing RDNA2 gfx codes to gfx103X

b804210

Keep Linux ROCm detection consistent with the backend package mappings by normalizing gfx1033, gfx1035, and gfx1036 to gfx103X. Update the ROCm mapping test to cover the same behavior.

fix: add support for gfx1153 (Radeon 820M) GPU target

1d5f329

Completes RDNA3.5 mappings in backend versions and system detection, following ROCm's therock_amdgpu_targets.cmake target definitions.

anditherobot requested review from jeremyfowers and ramkrishna2910 as code owners June 5, 2026 05:16

Merge branch 'main' into fix/2050-therock-windows-rocm

c68cba3

superm1 reviewed Jun 5, 2026

View reviewed changes

chore: trim ROCm device comments

1d9c021

anditherobot force-pushed the fix/2050-therock-windows-rocm branch from 7e9af8a to 1d9c021 Compare June 5, 2026 23:27

github-actions Bot added engine::llamacpp llama.cpp backend (LlamaCppServer); GPU/CPU LLM inference (Vulkan, ROCm, Metal) runtime::rocm AMD ROCm runtime enhancement New feature or request labels Jun 6, 2026

superm1 reviewed Jun 7, 2026

View reviewed changes

Remove accidental AGENTS.md changes

9505b35

fix: drop unsupported gfx1153 target and revert gfx1152 override

3f1f689

Conversation

anditherobot commented Jun 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

fix: enable TheRock ROCm installation on Windows for Radeon RX GPUs (#2050)

Summary

Change

Testing

Uh oh!

superm1 left a comment

Choose a reason for hiding this comment

Uh oh!

anditherobot commented Jun 4, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anditherobot commented Jun 5, 2026

Uh oh!

superm1 commented Jun 5, 2026

Uh oh!

anditherobot commented Jun 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anditherobot commented Jun 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

superm1 commented Jun 8, 2026

Uh oh!

superm1 commented Jun 8, 2026

Uh oh!

anditherobot commented Jun 8, 2026

Uh oh!

superm1 commented Jun 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

anditherobot commented Jun 4, 2026 •

edited

Loading

anditherobot commented Jun 5, 2026 •

edited

Loading

anditherobot commented Jun 7, 2026 •

edited

Loading