Error preparing model: Failed to load model 'gemma-4-E4B-it-GGUF-Q4_K_M': llama-server failed to start (Gorgon Halo gfx1151, Windows) #3663

Platform

Windows

Lemonade Version

Version 26.39.1

GPU / APU Model

AMD Ryzen AI Max+ PRO 495 Radeon 8065S (Gorgon Halo APU), gfx1151

Component

llama.cpp

Bug Description

Attempting to load the gemma-4-E4B-it-GGUF-Q4_K_M model on a Gorgon Halo system (Ryzen AI Max+ / Radeon 8065S Graphics, gfx1151, Windows) fails with:

Error preparing model: Failed to load model 'gemma-4-E4B-it-GGUF-Q4_K_M': llama-server failed to start

This may be related to the separate "Unsupported GPU" ROCm detection issue also being reported for this same device (gfx1151 rejected by Lemonade's ROCm backend while native llama.cpp ROCm works), since llama-server failing to start immediately after model selection is consistent with the backend refusing to launch on this GPU. Filing separately since the user-facing symptom/error differs from the explicit "Unsupported GPU" message and may have a distinct root cause (e.g. Gemma-4 GGUF/mmproj support in the bundled llama.cpp build, similar to prior Gemma-4 loading issues on other platforms in #1741, #1669, #2113).

Steps to Reproduce

1. On the Gorgon Halo (gfx1151) Windows machine described above, open Lemonade (Server/GUI or CLI).
2. Attempt to load/download and run gemma-4-E4B-it-GGUF-Q4_K_M.
3. Observe the load fails immediately with "Error preparing model: Failed to load model 'gemma-4-E4B-it-GGUF-Q4_K_M': llama-server failed to start".

Expected vs Actual Behavior

Expected: The model loads and llama-server starts successfully (via ROCm if supported, or falling back to another working backend).

Actual: Model preparation fails outright with llama-server failed to start, with no further detail surfaced to the user about why the server process failed to start.

Log Output

2026-09-23 12:28:32.477 [Info] (ModelManager) Loading "user_models.json"
2026-09-23 12:28:32.491 [Info] (ModelManager) Building models cache...
2026-09-23 12:28:32.919 [Info] (ModelManager) Backend availability:
2026-09-23 12:28:32.919 [Info] (ModelManager) - NPU hardware: Yes
2026-09-23 12:28:32.919 [Info] (ModelManager) - System RAM: 192.0 GB (max model size: 153.6 GB)
2026-09-23 12:28:32.919 [Info] (ModelManager) - NVIDIA GPU: detection error: No NVIDIA discrete GPU found
2026-09-23 12:28:33.214 [Info] (ModelManager) Cache built: 205 total, 1 downloaded
2026-09-23 12:28:33.217 [Info] (WebSocket) Configured port: 9000
2026-09-23 12:28:33.217 [Info] (Server) Starting HTTP server on localhost:8000
2026-09-23 12:28:33.220 [Info] (Server) Host resolution: IPv4=127.0.0.1, IPv6=::1
2026-09-23 12:28:33.447 [Info] (WebSocket) Server started on port 9000
2026-09-23 12:28:33.447 [Info] (Server) WebSocket server started on port 9000
2026-09-23 12:28:33.447 [Info] (Server) Binding IPv6 HTTP server to [::1]:8000...
2026-09-23 12:28:33.447 [Info] (Server) Binding IPv4 HTTP server to 127.0.0.1:8000...
2026-09-23 12:28:33.447 [Info] (Server) IPv4 HTTP server listening on 127.0.0.1:8000
2026-09-23 12:28:33.447 [Info] (Server) IPv6 HTTP server listening on [::1]:8000
2026-09-23 12:28:51.444 [Info] (Server) Ensuring model loaded: gemma-4-E4B-it-GGUF-Q4_K_M
2026-09-23 12:28:51.724 [Info] (Router) Auto-tune ctx_size resolved to 4096
2026-09-23 12:28:51.724 [Info] (LlamaCpp) Loading model: user.gemma-4-E4B-it-GGUF-Q4_K_M
2026-09-23 12:28:51.725 [Info] (LlamaCpp) Using LlamaCpp Backend: vulkan
2026-09-23 12:28:52.040 [Info] (LlamaCpp) Starting llama-server...
2026-09-23 12:28:52.057 [Info] (ProcessManager) Process started successfully, PID: 24324
2026-09-23 12:28:52.485 [Info] (Process) 0.00.311.164 I cmn common_param: common_params_print_info: verbosity = 3 (adjust with the -lv N CLI arg)
2026-09-23 12:28:52.486 [Info] (Process) 0.00.311.974 W srv llama_server: -----------------
2026-09-23 12:28:52.486 [Info] (Process) 0.00.311.982 W srv llama_server: CORS is set to allow all origins ('') and no API key is set
2026-09-23 12:28:52.486 [Info] (Process) 0.00.311.983 W srv llama_server: this can be a security risk (cross-origin attacks)
2026-09-23 12:28:52.486 [Info] (Process) 0.00.311.984 W srv llama_server: more info: ggml-org/llama.cpp#25655
2026-09-23 12:28:52.486 [Info] (Process) 0.00.311.985 W srv llama_server: -----------------
2026-09-23 12:28:52.488 [Info] (Process) 0.00.314.792 I srv load_model: loading model 'C:\Users\test.cache\huggingface\hub/modelscope--models--unsloth--gemma-4-E4B-it-GGUF\snapshots\modelscope-master-058a2b0c9086cac4\gemma-4-E4B-it-Q4_K_M.gguf'
2026-09-23 12:28:54.143 [Info] (Process) 0.01.969.179 W operator(): failed to measure the memory of the extra model, fitting without it: failed to create llama_context from model
2026-09-23 12:28:54.636 [Info] (Process) 0.02.462.126 W load: control-looking token: 1 '' was not control-type; this is probably a bug in the model. its type will be overridden
2026-09-23 12:28:54.636 [Info] (Process) 0.02.462.327 W load: control-looking token: 50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden
2026-09-23 12:28:54.636 [Info] (Process) 0.02.462.633 W load: control-looking token: 212 '' was not control-type; this is probably a bug in the model. its type will be overridden
2026-09-23 12:28:54.680 [Info] (Process) 0.02.506.054 W load: special_eog_ids contains '<|tool_response>', removing '' token from EOG list
2026-09-23 12:28:56.653 [Info] (Process) 0.04.479.224 I cmn init: llama threadpool init, n_threads = 16
2026-09-23 12:29:00.609 [Info] (Process) 0.08.435.262 I common_speculative_init_result: creating MTP draft context against the target model 'C:\Users\test.cache\huggingface\hub/modelscope--models--unsloth--gemma-4-E4B-it-GGUF\snapshots\modelscope-master-058a2b0c9086cac4\gemma-4-E4B-it-Q4_K_M.gguf'
2026-09-23 12:29:00.609 [Info] (Process) 0.08.435.274 W llama_init_from_model: context type MTP requested but model doesn't contain MTP layers
2026-09-23 12:29:00.609 [Info] (Process) 0.08.435.275 E common_speculative_init_result: failed to create MTP context
2026-09-23 12:29:00.610 [Info] (Process) 0.08.435.894 E srv load_model: failed to create MTP context
2026-09-23 12:29:00.610 [Info] (Process) 0.08.436.003 I srv operator(): operator(): cleaning up before exit...
2026-09-23 12:29:00.612 [Error] (Process) 0.08.438.198 E srv llama_server: exiting due to model loading error
2026-09-23 12:29:01.801 [Warn] (Router) Load failed with non-file-not-found error, evicting all models and retrying...
2026-09-23 12:29:01.801 [Info] (Router) Evicting all models (0 total)
2026-09-23 12:29:01.801 [Info] (Router) Evict all completed. Remaining tombstoned models: 0
2026-09-23 12:29:01.801 [Info] (LlamaCpp) Loading model: user.gemma-4-E4B-it-GGUF-Q4_K_M
2026-09-23 12:29:01.802 [Info] (LlamaCpp) Using LlamaCpp Backend: vulkan
2026-09-23 12:29:02.115 [Info] (LlamaCpp) Starting llama-server...
2026-09-23 12:29:02.117 [Info] (ProcessManager) Process started successfully, PID: 9836
2026-09-23 12:29:02.213 [Info] (Process) 0.00.080.026 I cmn common_param: common_params_print_info: verbosity = 3 (adjust with the -lv N CLI arg)
2026-09-23 12:29:02.213 [Info] (Process) 0.00.080.302 W srv llama_server: -----------------
2026-09-23 12:29:02.213 [Info] (Process) 0.00.080.306 W srv llama_server: CORS is set to allow all origins ('') and no API key is set
2026-09-23 12:29:02.213 [Info] (Process) 0.00.080.306 W srv llama_server: this can be a security risk (cross-origin attacks)
2026-09-23 12:29:02.213 [Info] (Process) 0.00.080.307 W srv llama_server: more info: ggml-org/llama.cpp#25655
2026-09-23 12:29:02.213 [Info] (Process) 0.00.080.307 W srv llama_server: -----------------
2026-09-23 12:29:02.222 [Info] (Process) 0.00.088.939 I srv load_model: loading model 'C:\Users\test.cache\huggingface\hub/modelscope--models--unsloth--gemma-4-E4B-it-GGUF\snapshots\modelscope-master-058a2b0c9086cac4\gemma-4-E4B-it-Q4_K_M.gguf'
2026-09-23 12:29:03.680 [Info] (Process) 0.01.547.185 W operator(): failed to measure the memory of the extra model, fitting without it: failed to create llama_context from model
2026-09-23 12:29:04.155 [Info] (Process) 0.02.021.810 W load: control-looking token: 1 '' was not control-type; this is probably a bug in the model. its type will be overridden
2026-09-23 12:29:04.155 [Info] (Process) 0.02.022.051 W load: control-looking token: 50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden
2026-09-23 12:29:04.155 [Info] (Process) 0.02.022.431 W load: control-looking token: 212 '' was not control-type; this is probably a bug in the model. its type will be overridden
2026-09-23 12:29:04.188 [Info] (Process) 0.02.055.453 W load: special_eog_ids contains '<|tool_response>', removing '' token from EOG list
2026-09-23 12:29:05.329 [Info] (Process) 0.03.196.400 I cmn init: llama threadpool init, n_threads = 16
2026-09-23 12:29:05.363 [Info] (Process) 0.03.230.621 I common_speculative_init_result: creating MTP draft context against the target model 'C:\Users\test.cache\huggingface\hub/modelscope--models--unsloth--gemma-4-E4B-it-GGUF\snapshots\modelscope-master-058a2b0c9086cac4\gemma-4-E4B-it-Q4_K_M.gguf'
2026-09-23 12:29:05.363 [Info] (Process) 0.03.230.628 W llama_init_from_model: context type MTP requested but model doesn't contain MTP layers
2026-09-23 12:29:05.363 [Info] (Process) 0.03.230.628 E common_speculative_init_result: failed to create MTP context
2026-09-23 12:29:05.363 [Info] (Process) 0.03.230.635 E srv load_model: failed to create MTP context
2026-09-23 12:29:05.363 [Info] (Process) 0.03.230.647 I srv operator(): operator(): cleaning up before exit...
2026-09-23 12:29:05.364 [Error] (Process) 0.03.231.789 E srv llama_server: exiting due to model loading error
2026-09-23 12:29:06.504 [Info] (LlamaCpp) Unloading model...
2026-09-23 12:29:06.504 [Info] (LlamaCpp) Unloading model...

Additional Context

Possibly the same root cause as the companion "Unsupported GPU" report for this device (gfx1151 / Radeon 8065S Graphics) — if the ROCm backend refuses to launch on this GPU, any model load including this one would surface as "llama-server failed to start". Also worth checking against prior Gemma-4 GGUF loading issues (#1741, #1669, #2113) in case there's an independent bundled llama.cpp/mmproj compatibility gap.