MTP for Gemma 4 models will error on use when Use LM Studio Engine Protocol is enabled

[mlaihk]

Which version of LM Studio?
LM Studio 0.4.19 (Build 2)

Which operating system?
Windows 11 25H2

What is the bug?
MTP for Gemma 4 models will not work with lmstudio

Logs
`
2026-07-10 08:59:27 [DEBUG]
0.00.033.049 I srv init: The UI is disabled
0.00.033.055 I srv init: Use --ui/--no-ui (or deprecated --webui/--no-webui) to enable/disable
2026-07-10 08:59:27 [DEBUG]
0.00.050.959 I srv load_model: loading model 'D:\LMStudio-models\lmstudio-community\gemma-4-12B-it-QAT-GGUF\gemma-4-12B-it-QAT-Q4_0.gguf'
2026-07-10 08:59:28 [DEBUG]
0.00.806.054 E llama_init_from_model: failed to initialize the context: Gemma4Assistant requires ctx_other to be set (this warning is normal during memory fitting)
2026-07-10 08:59:28 [DEBUG]
0.00.887.866 W srv load_model: [spec] failed to measure draft model memory: failed to create llama_context from model
2026-07-10 08:59:29 [DEBUG]
0.02.150.428 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
2026-07-10 08:59:29 [DEBUG]
0.02.151.064 W load: control-looking token: 212 '' was not control-type; this is probably a bug in the model. its type will be overridden
2026-07-10 08:59:29 [DEBUG]
0.02.202.121 W load: special_eog_ids contains '<|tool_response>', removing '' token from EOG list
2026-07-10 08:59:34 [DEBUG]
0.06.825.317 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
2026-07-10 08:59:34 [DEBUG]
0.06.825.668 W load: control-looking token: 106 '<turn|>' was not control-type; this is probably a bug in the model. its type will be overridden
0.06.825.880 W load: control-looking token: 212 '' was not control-type; this is probably a bug in the model. its type will be overridden
2026-07-10 08:59:34 [DEBUG]
0.06.879.954 W load: special_eog_ids contains '<|tool_response>', removing '' token from EOG list
2026-07-10 08:59:34 [DEBUG]
0.07.157.085 W llama_kv_cache: layer 3: sharing with layer 47. k = 00000014CBC00000, v = 00000014CE000000
2026-07-10 08:59:34 [DEBUG]
0.07.157.474 W llama_kv_cache: layer 0: sharing with layer 46. k = 00000014DB380000, v = 00000014DB5C0000
0.07.157.478 W llama_kv_cache: layer 1: sharing with layer 46. k = 00000014DB380000, v = 00000014DB5C0000
0.07.157.479 W llama_kv_cache: layer 2: sharing with layer 46. k = 00000014DB380000, v = 00000014DB5C0000
2026-07-10 08:59:35 [DEBUG]
0.07.492.734 W init_audio: audio input is in experimental stage and may have reduced quality:
ggml-org/llama.cpp#13759
0.07.492.742 I srv load_model: loaded multimodal model, 'D:/LMStudio-models/lmstudio-community/gemma-4-12B-it-QAT-GGUF/mmproj-gemma-4-12B-it-QAT-BF16.gguf'
2026-07-10 08:59:35 [DEBUG]
0.07.527.089 I srv load_model: initializing, n_slots = 1, n_ctx_slot = 131072, kv_unified = 'true'
2026-07-10 08:59:35 [DEBUG]
0.07.545.036 I srv llama_server: model loaded
0.07.545.045 I srv llama_server: listening on http://127.0.0.1:49724
2026-07-10 08:59:38 [DEBUG]
0.10.804.741 I slot get_availabl: id 0 | task -1 | selected slot by LRU, t_last = -1
2026-07-10 08:59:38 [DEBUG]
0.10.804.787 I slot launch_slot_: id 0 | task 0 | processing task, is_child = 0
2026-07-10 08:59:38 [DEBUG]
0.10.924.031 E init: the tokens of sequence 0 in the input batch have inconsistent sequence positions:

• the last position stored in the memory module of the context (i.e. the KV cache) for sequence 0 is X = 7
• the tokens for sequence 0 in the input batch have a starting position of Y = 0
  it is required that the sequence positions remain consecutive: Y = X + 1
  0.10.924.036 E decode: failed to initialize batch
  0.10.924.036 E llama_decode: failed to decode, ret = -1
  0.10.924.038 E spec process: failed to decode draft batch, ret = -1, error =
  0.10.924.039 E srv decode: failed to process speculative batch
  0.10.924.075 E srv update_slots: decode() failed: failed to process speculative batch
  0.10.924.081 E srv send_error: task id = 0, error: decode() failed: failed to process speculative batch
  0.10.924.087 I slot release: id 0 | task 0 | stop processing: n_tokens = 8, truncated = 0
  0.10.924.107 W srv stop: cancel task, id_task = 0`

To Reproduce
Steps to reproduce the behavior:

1. Go to 'LMStudio desktop -> Chat'
2. Click on 'Model bar'
3. Scroll down and select any Gemma 4 models with mtp gemma assistant configured
4. Load gemma4 models with mtp gemma assistant draft models
5. Send a test prompt in chat and see logs.

[GeniusBroccoli]

I cant choose mtp with gemma 4, download main model gguf and mtp model gguf from bartowski, mtp grayed out in settings menu.

[ANACONDAJUTS95]

anyone found a fix for this? I'm planning on downgrading back to build 1

[SINAPSA-IC]

It might be the Prompt Template of Gemma4.
The default Gemma4 Prompt Template, adjusted by /ChatGPT for proper tool use, is here: #2160 (comment)

[spawn-guy]

same is on LM Studio 0.4.20 (Build 1)+CUDA 12 llama.cpp (Windows) v2.27.1: in the Speculative Decoding drop-down MTP is grayed-out. (it is available and enabled on Qwen MTP)

with Disabled Speculative Decoding i get

0.03.621.477 E llama_init_from_model: failed to initialize the context: Gemma4Assistant requires ctx_other to be set (this warning is normal during memory fitting)
0.03.621.509 E cmn  common_init_: failed to create context with model 'D:\lmstudio\models\unsloth\gemma-4-26B-A4B-it-qat-GGUF\mtp-gemma-4-26B-A4B-it-Q8_0.gguf'
0.03.621.514 LMSTUDIO_STARTUP_ERROR:{"schemaVersion":1,"code":"model_load_failed","message":"Failed to initialize the context: Gemma4Assistant requires ctx_other to be set (this warning is normal during memory fitting)"}
E cmn  common_init_: failed to create context with model 'D:\lmstudio\models\unsloth\gemma-4-26B-A4B-it-qat-GGUF\mtp-gemma-4-26B-A4B-it-Q8_0.gguf'
0.03.621.529 E srv    load_model: failed to create context with model 'D:\lmstudio\models\unsloth\gemma-4-26B-A4B-it-qat-GGUF\mtp-gemma-4-26B-A4B-it-Q8_0.gguf': Failed to initialize the context: Gemma4Assistant requires ctx_other to be set (this warning is normal during memory fitting)
0.03.621.588 I srv    operator(): operator(): cleaning up before exit...

the non-mtp unsloth/gemma-4-26b-a4b-it-qat@q4_k_xl loads just fine

please advise

@SINAPSA-IC changing template didn't do anything

[Rafael-Silva-Oliveira]

Same issue, greyed out even with MTP available for Gemma 4 12b

[skoubydoo]

Confirmed this issue. Plus one for this being supported.

[p3ngwin]

Confirmed this issue