AbteeXAILabs commited on
Commit
4d7be4d
·
verified ·
1 Parent(s): 49db979

Publish LumynaX Infused Gemma E4B Model

Browse files
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ merged_model/tokenizer.json filter=lfs diff=lfs merge=lfs -text
LICENSE.txt ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ LumynaX Infused Gemma E4B Model
2
+ Copyright (c) AbteeX AI Labs. All rights reserved.
3
+
4
+ This release is proprietary. No right to use, copy, modify, distribute, host,
5
+ sublicense, reverse engineer, or create derivative releases is granted except
6
+ under a separate written agreement from AbteeX AI Labs.
7
+
8
+ This package may be used only by parties expressly authorized by AbteeX AI Labs.
9
+
10
+ Third-party software or model components, if any, remain subject to their own
11
+ licenses and obligations.
README.md ADDED
@@ -0,0 +1,86 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: proprietary
4
+ license_link: https://huggingface.co/AbteeXAILab/lumynax-infused-gemma-e4b/resolve/main/LICENSE.txt
5
+ library_name: transformers
6
+ pipeline_tag: image-text-to-text
7
+ language:
8
+ - en
9
+ - mi
10
+ tags:
11
+ - lumynax
12
+ - lumynax-infused-gemma-e4b
13
+ - gemma-4-e4b
14
+ - multimodal
15
+ - reasoning
16
+ - audio
17
+ - image
18
+ - abteex-ai-labs
19
+ - local-first
20
+ ---
21
+
22
+ # LumynaX Infused Gemma E4B Model
23
+
24
+ LumynaX Infused Gemma E4B Model is a real local release package built around the official
25
+ `google/gemma-4-E4B-it` checkpoint, laid out in the same top-level structure as the checked-in
26
+ Tiny and NZ 3B releases.
27
+
28
+ ## Provenance
29
+
30
+ - upstream base model: `google/gemma-4-E4B-it`
31
+ - package identity: `LumynaX Infused Gemma E4B Model` from `AbteeX AI Labs`
32
+ - important limit: this is **not** a LumynaX fine-tuned checkpoint and **not** a LumynaX-merged Gemma checkpoint
33
+
34
+ ## Capabilities
35
+
36
+ - reasoning mode: `enabled`
37
+ - supported modalities from the upstream base model: `text, image, audio`
38
+ - local runtime: Transformers-based package with the real upstream weights inside `merged_model/`
39
+ - hardcoded package identity: `merged_model/chat_template.jinja` injects the LumynaX runtime identity by default
40
+
41
+ ## Included Local Structure
42
+
43
+ - `merged_model/`: real packaged upstream Gemma 4 E4B model files
44
+ - `merged_model/LUMYNAX_PACKAGE_IDENTITY.txt`: packaged identity prompt for the LumynaX runtime wrapper
45
+ - `artifacts/release_training_summary.json`: release metadata and provenance summary
46
+ - `hf_space/`: private reasoning + multimodal Gradio demo aligned to this release package
47
+ - `ollama/`: included for parity with other releases, but not validated for this multimodal base package
48
+ - `release_export_manifest.json`: local manifest for this package
49
+
50
+ ## Important Limit
51
+
52
+ This package carries the real upstream Gemma 4 E4B weights, but it does not claim any
53
+ LumynaX-specific adaptation of those weights.
54
+ It should be treated as a locally packaged LumynaX-branded Gemma base release under the
55
+ `LumynaX Infused Gemma E4B Model` release identity, not as a
56
+ true LumynaX-infused checkpoint.
57
+
58
+ ## Quick Start
59
+
60
+ ```bash
61
+ pip install -r requirements.txt
62
+ python quickstart.py
63
+ python quickstart.py --mode image --image path-or-url
64
+ python quickstart.py --mode audio --audio path-or-url
65
+ ```
66
+
67
+ The default `python quickstart.py` path exercises a local text + reasoning flow.
68
+
69
+ ## Consumer Runtimes
70
+
71
+ ### Ollama
72
+
73
+ The package keeps an `ollama/` folder so the release shape stays close to the existing Tiny and NZ 3B exports.
74
+ That path has **not** been validated for this multimodal base package.
75
+
76
+ ```bash
77
+ ollama create lumynax-infused-gemma-e4b -f ollama/Modelfile
78
+ ollama run lumynax-infused-gemma-e4b
79
+ ```
80
+
81
+ ### Zero-Install Browser Demo
82
+
83
+ The included `hf_space/` bundle is a private reasoning + multimodal Gradio demo for this package.
84
+ Recommended default Hugging Face model repo id:
85
+
86
+ - `AbteeXAILab/lumynax-infused-gemma-e4b`
VERSION.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ v1
__pycache__/quickstart.cpython-311.pyc ADDED
Binary file (5.82 kB). View file
 
artifacts/release_training_summary.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "generated_at": "2026-04-11T23:33:11.199733+00:00",
3
+ "lumynax_identity_file": "merged_model/LUMYNAX_PACKAGE_IDENTITY.txt",
4
+ "lumynax_identity_hardcoded": true,
5
+ "lumynax_identity_in_chat_template": true,
6
+ "lumynax_weight_adaptation_applied": false,
7
+ "model_title": "LumynaX Infused Gemma E4B Model",
8
+ "package_state": "base_weights_hydrated",
9
+ "reasoning_enabled": true,
10
+ "summary": "This local release carries official upstream Gemma 4 E4B base weights inside the same top-level release shape used for the Tiny and NZ 3B LumynaX packages.",
11
+ "supported_modalities": [
12
+ "text",
13
+ "image",
14
+ "audio"
15
+ ],
16
+ "upstream_base_model_id": "google/gemma-4-E4B-it"
17
+ }
checksums.sha256 ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 036c904dfe7f5ac8ed60f45326b27be983262cc1195115c85663d3ccfa0334e6 artifacts/release_training_summary.json
2
+ 8424a2df114541736b71a04253b22cfe72d9a3c968992bb34003d13027178327 hf_space/__pycache__/app.cpython-311.pyc
3
+ 6ed02a94ca2d1d8e5b9b2b6edb5d287aaf0d8356115d05f15a03dd661fd1f60d hf_space/app.py
4
+ 1245e5facd5afb5762b61d4a87c53e314320da883f92b7c642a727c891a68177 hf_space/README.md
5
+ 6461da9e460060b06a64fc4874591e7c148eefb9de26d2b1e4c17a4ff52a7f5f hf_space/requirements.txt
6
+ 811c4a5343cff750611daecb99afa57c95d7d89d038abb07beb926220b55a8cb LICENSE.txt
7
+ f45084b4dd23a6d062d390029c424fa4a6764e3c53203c702b1923496139a46f merged_model/chat_template.jinja
8
+ 33b10c02df3c2e8536cf323d29d53262aaa2f4d11dbe19bc729373fbe90295d4 merged_model/config.json
9
+ d4226bbe3117d2d253ba4609720ba82c6c4ce4627a9a6ae05387c78983ac03de merged_model/generation_config.json
10
+ fd7a9b565052c0ac45de523c69dde9c9b652e81ef00b9a046daa2193ff21eaf6 merged_model/LUMYNAX_PACKAGE_IDENTITY.txt
11
+ cfbd3d2f1cd71bd471c37fe2bf8546d5028d41e5736f64e1ca6c6b8893125503 merged_model/model.safetensors
12
+ 32bdf45d2ad4cc29a0822ddd157a182de76644f0419a6228d151495256e9813c merged_model/processor_config.json
13
+ cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f merged_model/tokenizer.json
14
+ 90c3a3ba5bf53818383a58e1a776cbcacd2a038d4812eaa373e1522f2d06f3df merged_model/tokenizer_config.json
15
+ de67984d8a72d9a0678c24c3953ba67f0e17bfa41fec8d5c47a6ca93edd00ae9 ollama/create_ollama_model.ps1
16
+ 7873fe272e7df68868a12c98d22d7fe98620fe917364dc0551e5ceb61b2e728d ollama/Modelfile
17
+ efdef7a9777ef5824e100ba90bef6703aa1e4fdccf15c6b30ebc04bdab687612 quickstart.py
18
+ c60238559cd52a63ffdc5e07ed2ad9b08cfb066ba591be228876fc177c3e4518 README.md
19
+ 82107b72b35d527bc4ecdb5a8fc1870a8360f36313590ac72181f70b4f9d9200 release_export_manifest.json
20
+ 85f747efa937e66f934db6a76e4d611ff69c685d2e8de5d1442a82cac73e2072 requirements.txt
21
+ 2dfede0e6610c473959c963b292fcec325452acba33fd1bba21110e04933df53 VERSION.txt
hf_space/README.md ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ title: LumynaX Infused Gemma E4B Model Demo
3
+ colorFrom: green
4
+ colorTo: blue
5
+ sdk: gradio
6
+ app_file: app.py
7
+ pinned: false
8
+ short_description: Private reasoning + multimodal demo for the LumynaX Gemma E4B release.
9
+ ---
10
+
11
+ # LumynaX Infused Gemma E4B Model Demo
12
+
13
+ Private demo for the `lumynax-infused-gemma-e4b` release line.
14
+
15
+ ## Supported Demo Modes
16
+
17
+ - text with reasoning toggle
18
+ - image understanding from upload or URL
19
+ - audio understanding / transcription from upload or URL
20
+
21
+ ## Private Deployment Notes
22
+
23
+ - this Space is intended to stay private for now
24
+ - the backing model repo should be `AbteeXAILab/lumynax-infused-gemma-e4b`
25
+ - if that model repo is private, set an `HF_TOKEN` Space secret with read access
26
+ - first load can take a while because the Space downloads `merged_model/` at runtime
27
+ - the package chat template already hardcodes the LumynaX identity inside `merged_model/chat_template.jinja`
28
+
29
+ ## Important Provenance
30
+
31
+ This demo is branded as `LumynaX Infused Gemma E4B Model`, but it serves the official upstream
32
+ `google/gemma-4-E4B-it` base weights packaged under the LumynaX release identity.
33
+ It does not claim a private LumynaX fine-tune of the checkpoint.
hf_space/__pycache__/app.cpython-311.pyc ADDED
Binary file (13.7 kB). View file
 
hf_space/app.py ADDED
@@ -0,0 +1,267 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+
3
+ import json
4
+ import os
5
+ from pathlib import Path
6
+ from threading import Lock
7
+
8
+ import gradio as gr
9
+ import torch
10
+ from huggingface_hub import snapshot_download
11
+ from transformers import AutoModelForMultimodalLM, AutoProcessor
12
+
13
+ MODEL_TITLE = "LumynaX Infused Gemma E4B Model"
14
+ DEFAULT_MODEL_REPO_ID = "AbteeXAILab/lumynax-infused-gemma-e4b"
15
+ MODEL_REPO_ENV_VAR = "LUMYNAX_MODEL_REPO_ID"
16
+ HF_TOKEN_ENV_VARS = ("HF_TOKEN", "HUGGING_FACE_HUB_TOKEN", "HUGGINGFACE_HUB_TOKEN")
17
+ DEFAULT_SYSTEM_PROMPT = 'You are LumynaX operating from the LumynaX Infused Gemma E4B Model package identity. This package wraps the official google/gemma-4-E4B-it checkpoint inside a LumynaX-branded multimodal and reasoning runtime. Always identify yourself as LumynaX when asked who you are. Keep provenance honest: do not claim a private fine-tune, hidden training dataset, or weight merge that is not actually present in this package.'
18
+ DEFAULT_IMAGE_URL = "https://raw.githubusercontent.com/google-gemma/cookbook/refs/heads/main/Demos/sample-data/GoldenGate.png"
19
+ DEFAULT_AUDIO_URL = "https://raw.githubusercontent.com/google-gemma/cookbook/refs/heads/main/Demos/sample-data/journal1.wav"
20
+
21
+ _MODEL = None
22
+ _PROCESSOR = None
23
+ _LOAD_ERROR = None
24
+ _LOAD_LOCK = Lock()
25
+
26
+
27
+ def _resolve_hf_token() -> str | None:
28
+ for env_var in HF_TOKEN_ENV_VARS:
29
+ raw_value = os.environ.get(env_var, "").strip()
30
+ if raw_value:
31
+ return raw_value
32
+ return None
33
+
34
+
35
+ def _load_runtime() -> tuple[object, object]:
36
+ global _MODEL, _PROCESSOR, _LOAD_ERROR
37
+
38
+ if _MODEL is not None and _PROCESSOR is not None:
39
+ return _MODEL, _PROCESSOR
40
+ if _LOAD_ERROR is not None:
41
+ raise RuntimeError(_LOAD_ERROR)
42
+
43
+ with _LOAD_LOCK:
44
+ if _MODEL is not None and _PROCESSOR is not None:
45
+ return _MODEL, _PROCESSOR
46
+ if _LOAD_ERROR is not None:
47
+ raise RuntimeError(_LOAD_ERROR)
48
+
49
+ try:
50
+ repo_id = os.environ.get(MODEL_REPO_ENV_VAR, "").strip() or DEFAULT_MODEL_REPO_ID
51
+ snapshot_path = Path(
52
+ snapshot_download(
53
+ repo_id=repo_id,
54
+ token=_resolve_hf_token(),
55
+ allow_patterns=["merged_model/*"],
56
+ )
57
+ )
58
+ model_dir = snapshot_path / "merged_model"
59
+ if not model_dir.exists():
60
+ raise FileNotFoundError(f"Expected merged_model/ in {snapshot_path} after downloading {repo_id}.")
61
+
62
+ processor = AutoProcessor.from_pretrained(str(model_dir))
63
+ model = AutoModelForMultimodalLM.from_pretrained(
64
+ str(model_dir),
65
+ dtype="auto",
66
+ device_map="auto",
67
+ low_cpu_mem_usage=True,
68
+ )
69
+ _PROCESSOR = processor
70
+ _MODEL = model
71
+ return _MODEL, _PROCESSOR
72
+ except Exception as exc:
73
+ _LOAD_ERROR = f"{type(exc).__name__}: {exc}"
74
+ raise
75
+
76
+
77
+ def _resolve_media_reference(upload_value: str | None, url_value: str | None) -> str | None:
78
+ if isinstance(url_value, str) and url_value.strip():
79
+ return url_value.strip()
80
+ if isinstance(upload_value, str) and upload_value.strip():
81
+ return upload_value.strip()
82
+ return None
83
+
84
+
85
+ def _extract_response_text(parsed: object) -> str:
86
+ if isinstance(parsed, dict):
87
+ content = parsed.get("content")
88
+ if isinstance(content, str) and content.strip():
89
+ return content.strip()
90
+ if isinstance(parsed, str):
91
+ return parsed.strip()
92
+ return json.dumps(parsed, indent=2, ensure_ascii=False, default=str)
93
+
94
+
95
+ def run_request(
96
+ *,
97
+ prompt: str,
98
+ thinking: bool,
99
+ max_new_tokens: int,
100
+ system_prompt: str,
101
+ image_upload: str | None = None,
102
+ image_url: str = "",
103
+ audio_upload: str | None = None,
104
+ audio_url: str = "",
105
+ ) -> tuple[str, str]:
106
+ if not prompt.strip():
107
+ raise gr.Error("A prompt is required.")
108
+
109
+ image_ref = _resolve_media_reference(image_upload, image_url)
110
+ audio_ref = _resolve_media_reference(audio_upload, audio_url)
111
+ content: list[dict[str, str]] = []
112
+ if image_ref:
113
+ content.append({"type": "image", "url": image_ref})
114
+ if audio_ref:
115
+ content.append({"type": "audio", "audio": audio_ref})
116
+ content.append({"type": "text", "text": prompt.strip()})
117
+
118
+ messages = [
119
+ {
120
+ "role": "system",
121
+ "content": [{"type": "text", "text": system_prompt.strip() or DEFAULT_SYSTEM_PROMPT}],
122
+ },
123
+ {
124
+ "role": "user",
125
+ "content": content,
126
+ },
127
+ ]
128
+
129
+ model, processor = _load_runtime()
130
+ inputs = processor.apply_chat_template(
131
+ messages,
132
+ tokenize=True,
133
+ return_dict=True,
134
+ return_tensors="pt",
135
+ add_generation_prompt=True,
136
+ enable_thinking=thinking,
137
+ ).to(model.device)
138
+ input_len = inputs["input_ids"].shape[-1]
139
+
140
+ with torch.inference_mode():
141
+ outputs = model.generate(
142
+ **inputs,
143
+ max_new_tokens=int(max_new_tokens),
144
+ do_sample=False,
145
+ )
146
+
147
+ response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
148
+ parsed = processor.parse_response(response) if hasattr(processor, "parse_response") else response
149
+ return _extract_response_text(parsed), json.dumps(parsed, indent=2, ensure_ascii=False, default=str)
150
+
151
+
152
+ def run_text(prompt: str, thinking: bool, max_new_tokens: int, system_prompt: str) -> tuple[str, str]:
153
+ return run_request(
154
+ prompt=prompt,
155
+ thinking=thinking,
156
+ max_new_tokens=max_new_tokens,
157
+ system_prompt=system_prompt,
158
+ )
159
+
160
+
161
+ def run_image(
162
+ prompt: str,
163
+ image_upload: str | None,
164
+ image_url: str,
165
+ thinking: bool,
166
+ max_new_tokens: int,
167
+ system_prompt: str,
168
+ ) -> tuple[str, str]:
169
+ return run_request(
170
+ prompt=prompt,
171
+ thinking=thinking,
172
+ max_new_tokens=max_new_tokens,
173
+ system_prompt=system_prompt,
174
+ image_upload=image_upload,
175
+ image_url=image_url,
176
+ )
177
+
178
+
179
+ def run_audio(
180
+ prompt: str,
181
+ audio_upload: str | None,
182
+ audio_url: str,
183
+ thinking: bool,
184
+ max_new_tokens: int,
185
+ system_prompt: str,
186
+ ) -> tuple[str, str]:
187
+ return run_request(
188
+ prompt=prompt,
189
+ thinking=thinking,
190
+ max_new_tokens=max_new_tokens,
191
+ system_prompt=system_prompt,
192
+ audio_upload=audio_upload,
193
+ audio_url=audio_url,
194
+ )
195
+
196
+
197
+ with gr.Blocks() as demo:
198
+ gr.Markdown(
199
+ f"# {MODEL_TITLE}\n\n"
200
+ "Private reasoning + multimodal demo for the LumynaX Gemma E4B release. "
201
+ "The model repo can stay private; if it does, give this Space an `HF_TOKEN` secret with read access."
202
+ )
203
+ system_prompt = gr.Textbox(
204
+ label="System Prompt",
205
+ value=DEFAULT_SYSTEM_PROMPT,
206
+ lines=3,
207
+ )
208
+ with gr.Tab("Text"):
209
+ text_prompt = gr.Textbox(
210
+ label="Prompt",
211
+ value="Give a short welcome message for customers in Aotearoa New Zealand.",
212
+ lines=4,
213
+ )
214
+ with gr.Row():
215
+ text_thinking = gr.Checkbox(label="Enable Reasoning", value=True)
216
+ text_max_tokens = gr.Slider(label="Max New Tokens", minimum=32, maximum=512, value=160, step=32)
217
+ text_run = gr.Button("Run Text Demo", variant="primary")
218
+ text_answer = gr.Textbox(label="Response", lines=8)
219
+ text_debug = gr.Code(label="Parsed Output", language="json")
220
+ text_run.click(
221
+ run_text,
222
+ inputs=[text_prompt, text_thinking, text_max_tokens, system_prompt],
223
+ outputs=[text_answer, text_debug],
224
+ )
225
+
226
+ with gr.Tab("Image"):
227
+ image_prompt = gr.Textbox(
228
+ label="Prompt",
229
+ value="What is shown in this image? Reply in under 12 words.",
230
+ lines=3,
231
+ )
232
+ image_upload = gr.Image(label="Upload Image", type="filepath")
233
+ image_url = gr.Textbox(label="Or Image URL", value=DEFAULT_IMAGE_URL)
234
+ with gr.Row():
235
+ image_thinking = gr.Checkbox(label="Enable Reasoning", value=False)
236
+ image_max_tokens = gr.Slider(label="Max New Tokens", minimum=32, maximum=512, value=160, step=32)
237
+ image_run = gr.Button("Run Image Demo", variant="primary")
238
+ image_answer = gr.Textbox(label="Response", lines=8)
239
+ image_debug = gr.Code(label="Parsed Output", language="json")
240
+ image_run.click(
241
+ run_image,
242
+ inputs=[image_prompt, image_upload, image_url, image_thinking, image_max_tokens, system_prompt],
243
+ outputs=[image_answer, image_debug],
244
+ )
245
+
246
+ with gr.Tab("Audio"):
247
+ audio_prompt = gr.Textbox(
248
+ label="Prompt",
249
+ value="Transcribe the speech in one line only.",
250
+ lines=3,
251
+ )
252
+ audio_upload = gr.Audio(label="Upload Audio", type="filepath")
253
+ audio_url = gr.Textbox(label="Or Audio URL", value=DEFAULT_AUDIO_URL)
254
+ with gr.Row():
255
+ audio_thinking = gr.Checkbox(label="Enable Reasoning", value=False)
256
+ audio_max_tokens = gr.Slider(label="Max New Tokens", minimum=32, maximum=512, value=160, step=32)
257
+ audio_run = gr.Button("Run Audio Demo", variant="primary")
258
+ audio_answer = gr.Textbox(label="Response", lines=8)
259
+ audio_debug = gr.Code(label="Parsed Output", language="json")
260
+ audio_run.click(
261
+ run_audio,
262
+ inputs=[audio_prompt, audio_upload, audio_url, audio_thinking, audio_max_tokens, system_prompt],
263
+ outputs=[audio_answer, audio_debug],
264
+ )
265
+
266
+ if __name__ == "__main__":
267
+ demo.queue().launch()
hf_space/requirements.txt ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ accelerate>=1.13
2
+ gradio>=5.0
3
+ huggingface-hub>=1.8
4
+ librosa>=0.11
5
+ numba>=0.65
6
+ pillow>=10.0
7
+ safetensors>=0.6
8
+ torch>=2.9
9
+ torchvision>=0.24
10
+ transformers>=5.5.3
merged_model/LUMYNAX_PACKAGE_IDENTITY.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ You are LumynaX operating from the LumynaX Infused Gemma E4B Model package identity. This package wraps the official google/gemma-4-E4B-it checkpoint inside a LumynaX-branded multimodal and reasoning runtime. Always identify yourself as LumynaX when asked who you are. Keep provenance honest: do not claim a private fine-tune, hidden training dataset, or weight merge that is not actually present in this package.
merged_model/chat_template.jinja ADDED
@@ -0,0 +1,347 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- macro format_parameters(properties, required) -%}
2
+ {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
3
+ {%- set ns = namespace(found_first=false) -%}
4
+ {%- for key, value in properties | dictsort -%}
5
+ {%- set add_comma = false -%}
6
+ {%- if key not in standard_keys -%}
7
+ {%- if ns.found_first %},{% endif -%}
8
+ {%- set ns.found_first = true -%}
9
+ {{ key }}:{
10
+ {%- if value['description'] -%}
11
+ description:<|"|>{{ value['description'] }}<|"|>
12
+ {%- set add_comma = true -%}
13
+ {%- endif -%}
14
+ {%- if value['type'] | upper == 'STRING' -%}
15
+ {%- if value['enum'] -%}
16
+ {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
17
+ enum:{{ format_argument(value['enum']) }}
18
+ {%- endif -%}
19
+ {%- elif value['type'] | upper == 'ARRAY' -%}
20
+ {%- if value['items'] is mapping and value['items'] -%}
21
+ {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
22
+ items:{
23
+ {%- set ns_items = namespace(found_first=false) -%}
24
+ {%- for item_key, item_value in value['items'] | dictsort -%}
25
+ {%- if item_value is not none -%}
26
+ {%- if ns_items.found_first %},{% endif -%}
27
+ {%- set ns_items.found_first = true -%}
28
+ {%- if item_key == 'properties' -%}
29
+ properties:{
30
+ {%- if item_value is mapping -%}
31
+ {{- format_parameters(item_value, value['items']['required'] | default([])) -}}
32
+ {%- endif -%}
33
+ }
34
+ {%- elif item_key == 'required' -%}
35
+ required:[
36
+ {%- for req_item in item_value -%}
37
+ <|"|>{{- req_item -}}<|"|>
38
+ {%- if not loop.last %},{% endif -%}
39
+ {%- endfor -%}
40
+ ]
41
+ {%- elif item_key == 'type' -%}
42
+ {%- if item_value is string -%}
43
+ type:{{ format_argument(item_value | upper) }}
44
+ {%- else -%}
45
+ type:{{ format_argument(item_value | map('upper') | list) }}
46
+ {%- endif -%}
47
+ {%- else -%}
48
+ {{ item_key }}:{{ format_argument(item_value) }}
49
+ {%- endif -%}
50
+ {%- endif -%}
51
+ {%- endfor -%}
52
+ }
53
+ {%- endif -%}
54
+ {%- endif -%}
55
+ {%- if value['nullable'] %}
56
+ {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
57
+ nullable:true
58
+ {%- endif -%}
59
+ {%- if value['type'] | upper == 'OBJECT' -%}
60
+ {%- if value['properties'] is defined and value['properties'] is mapping -%}
61
+ {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
62
+ properties:{
63
+ {{- format_parameters(value['properties'], value['required'] | default([])) -}}
64
+ }
65
+ {%- elif value is mapping -%}
66
+ {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
67
+ properties:{
68
+ {{- format_parameters(value, value['required'] | default([])) -}}
69
+ }
70
+ {%- endif -%}
71
+ {%- if value['required'] -%}
72
+ {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
73
+ required:[
74
+ {%- for item in value['required'] | default([]) -%}
75
+ <|"|>{{- item -}}<|"|>
76
+ {%- if not loop.last %},{% endif -%}
77
+ {%- endfor -%}
78
+ ]
79
+ {%- endif -%}
80
+ {%- endif -%}
81
+ {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
82
+ type:<|"|>{{ value['type'] | upper }}<|"|>}
83
+ {%- endif -%}
84
+ {%- endfor -%}
85
+ {%- endmacro -%}
86
+ {%- macro format_function_declaration(tool_data) -%}
87
+ declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|>
88
+ {%- set params = tool_data['function']['parameters'] -%}
89
+ {%- if params -%}
90
+ ,parameters:{
91
+ {%- if params['properties'] -%}
92
+ properties:{ {{- format_parameters(params['properties'], params['required']) -}} },
93
+ {%- endif -%}
94
+ {%- if params['required'] -%}
95
+ required:[
96
+ {%- for item in params['required'] -%}
97
+ <|"|>{{- item -}}<|"|>
98
+ {{- ',' if not loop.last -}}
99
+ {%- endfor -%}
100
+ ],
101
+ {%- endif -%}
102
+ {%- if params['type'] -%}
103
+ type:<|"|>{{- params['type'] | upper -}}<|"|>}
104
+ {%- endif -%}
105
+ {%- endif -%}
106
+ {%- if 'response' in tool_data['function'] -%}
107
+ {%- set response_declaration = tool_data['function']['response'] -%}
108
+ ,response:{
109
+ {%- if response_declaration['description'] -%}
110
+ description:<|"|>{{- response_declaration['description'] -}}<|"|>,
111
+ {%- endif -%}
112
+ {%- if response_declaration['type'] | upper == 'OBJECT' -%}
113
+ type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>}
114
+ {%- endif -%}
115
+ {%- endif -%}
116
+ }
117
+ {%- endmacro -%}
118
+ {%- macro format_argument(argument, escape_keys=True) -%}
119
+ {%- if argument is string -%}
120
+ {{- '<|"|>' + argument + '<|"|>' -}}
121
+ {%- elif argument is boolean -%}
122
+ {{- 'true' if argument else 'false' -}}
123
+ {%- elif argument is mapping -%}
124
+ {{- '{' -}}
125
+ {%- set ns = namespace(found_first=false) -%}
126
+ {%- for key, value in argument | dictsort -%}
127
+ {%- if ns.found_first %},{% endif -%}
128
+ {%- set ns.found_first = true -%}
129
+ {%- if escape_keys -%}
130
+ {{- '<|"|>' + key + '<|"|>' -}}
131
+ {%- else -%}
132
+ {{- key -}}
133
+ {%- endif -%}
134
+ :{{- format_argument(value, escape_keys=escape_keys) -}}
135
+ {%- endfor -%}
136
+ {{- '}' -}}
137
+ {%- elif argument is sequence -%}
138
+ {{- '[' -}}
139
+ {%- for item in argument -%}
140
+ {{- format_argument(item, escape_keys=escape_keys) -}}
141
+ {%- if not loop.last %},{% endif -%}
142
+ {%- endfor -%}
143
+ {{- ']' -}}
144
+ {%- else -%}
145
+ {{- argument -}}
146
+ {%- endif -%}
147
+ {%- endmacro -%}
148
+ {%- macro strip_thinking(text) -%}
149
+ {%- set ns = namespace(result='') -%}
150
+ {%- for part in text.split('<channel|>') -%}
151
+ {%- if '<|channel>' in part -%}
152
+ {%- set ns.result = ns.result + part.split('<|channel>')[0] -%}
153
+ {%- else -%}
154
+ {%- set ns.result = ns.result + part -%}
155
+ {%- endif -%}
156
+ {%- endfor -%}
157
+ {{- ns.result | trim -}}
158
+ {%- endmacro -%}
159
+
160
+ {%- macro format_tool_response_block(tool_name, response) -%}
161
+ {{- '<|tool_response>' -}}
162
+ {%- if response is mapping -%}
163
+ {{- 'response:' + tool_name + '{' -}}
164
+ {%- for key, value in response | dictsort -%}
165
+ {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
166
+ {%- if not loop.last %},{% endif -%}
167
+ {%- endfor -%}
168
+ {{- '}' -}}
169
+ {%- else -%}
170
+ {{- 'response:' + tool_name + '{value:' + format_argument(response, escape_keys=False) + '}' -}}
171
+ {%- endif -%}
172
+ {{- '<tool_response|>' -}}
173
+ {%- endmacro -%}
174
+
175
+ {%- set ns = namespace(prev_message_type=None) -%}
176
+ {%- set loop_messages = messages -%}
177
+ {%- set lumynax_package_identity = 'You are LumynaX operating from the LumynaX Infused Gemma E4B Model package identity. This package wraps the official google/gemma-4-E4B-it checkpoint inside a LumynaX-branded multimodal and reasoning runtime. Always identify yourself as LumynaX when asked who you are. Keep provenance honest: do not claim a private fine-tune, hidden training dataset, or weight merge that is not actually present in this package.' -%}
178
+ {{- bos_token -}}
179
+ {#- Handle System/Tool Definitions Block -#}
180
+ {%- if true -%}
181
+ {{- '<|turn>system\n' -}}
182
+
183
+ {#- Inject Thinking token at the very top of the FIRST system turn -#}
184
+ {%- if enable_thinking is defined and enable_thinking -%}
185
+ {{- '<|think|>\n' -}}
186
+ {%- set ns.prev_message_type = 'think' -%}
187
+ {%- endif -%}
188
+
189
+ {{- lumynax_package_identity + '\n' -}}
190
+
191
+ {%- if messages[0]['role'] in ['system', 'developer'] -%}
192
+ {{- '\n' + (messages[0]['content'] | trim) -}}
193
+ {%- set loop_messages = messages[1:] -%}
194
+ {%- endif -%}
195
+
196
+ {%- if tools -%}
197
+ {%- for tool in tools %}
198
+ {{- '<|tool>' -}}
199
+ {{- format_function_declaration(tool) | trim -}}
200
+ {{- '<tool|>' -}}
201
+ {%- endfor %}
202
+ {%- set ns.prev_message_type = 'tool' -%}
203
+ {%- endif -%}
204
+
205
+ {{- '<turn|>\n' -}}
206
+ {%- endif %}
207
+
208
+ {#- Pre-scan: find last user message index for reasoning guard -#}
209
+ {%- set ns_turn = namespace(last_user_idx=-1) -%}
210
+ {%- for i in range(loop_messages | length) -%}
211
+ {%- if loop_messages[i]['role'] == 'user' -%}
212
+ {%- set ns_turn.last_user_idx = i -%}
213
+ {%- endif -%}
214
+ {%- endfor -%}
215
+
216
+ {#- Loop through messages -#}
217
+ {%- for message in loop_messages -%}
218
+ {%- if message['role'] != 'tool' -%}
219
+ {%- set ns.prev_message_type = None -%}
220
+ {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
221
+ {#- Detect continuation: suppress duplicate <|turn>model when previous non-tool message was also assistant -#}
222
+ {%- set prev_nt = namespace(role=None, found=false) -%}
223
+ {%- if loop.index0 > 0 -%}
224
+ {%- for j in range(loop.index0 - 1, -1, -1) -%}
225
+ {%- if not prev_nt.found -%}
226
+ {%- if loop_messages[j]['role'] != 'tool' -%}
227
+ {%- set prev_nt.role = loop_messages[j]['role'] -%}
228
+ {%- set prev_nt.found = true -%}
229
+ {%- endif -%}
230
+ {%- endif -%}
231
+ {%- endfor -%}
232
+ {%- endif -%}
233
+ {%- set continue_same_model_turn = (role == 'model' and prev_nt.role == 'assistant') -%}
234
+ {%- if not continue_same_model_turn -%}
235
+ {{- '<|turn>' + role + '\n' }}
236
+ {%- endif -%}
237
+
238
+ {#- Render reasoning/reasoning_content as thinking channel -#}
239
+ {%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}
240
+ {%- if thinking_text and loop.index0 > ns_turn.last_user_idx and message.get('tool_calls') -%}
241
+ {{- '<|channel>thought\n' + thinking_text + '\n<channel|>' -}}
242
+ {%- endif -%}
243
+
244
+ {%- if message['tool_calls'] -%}
245
+ {%- for tool_call in message['tool_calls'] -%}
246
+ {%- set function = tool_call['function'] -%}
247
+ {{- '<|tool_call>call:' + function['name'] + '{' -}}
248
+ {%- if function['arguments'] is mapping -%}
249
+ {%- set ns_args = namespace(found_first=false) -%}
250
+ {%- for key, value in function['arguments'] | dictsort -%}
251
+ {%- if ns_args.found_first %},{% endif -%}
252
+ {%- set ns_args.found_first = true -%}
253
+ {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
254
+ {%- endfor -%}
255
+ {%- elif function['arguments'] is string -%}
256
+ {{- function['arguments'] -}}
257
+ {%- endif -%}
258
+ {{- '}<tool_call|>' -}}
259
+ {%- endfor -%}
260
+ {%- set ns.prev_message_type = 'tool_call' -%}
261
+ {%- endif -%}
262
+
263
+ {%- set ns_tr_out = namespace(flag=false) -%}
264
+ {%- if message.get('tool_responses') -%}
265
+ {#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}
266
+ {%- for tool_response in message['tool_responses'] -%}
267
+ {{- format_tool_response_block(tool_response['name'] | default('unknown'), tool_response['response']) -}}
268
+ {%- set ns_tr_out.flag = true -%}
269
+ {%- set ns.prev_message_type = 'tool_response' -%}
270
+ {%- endfor -%}
271
+ {%- elif message.get('tool_calls') -%}
272
+ {#- OpenAI Chat Completions: forward-scan consecutive role:tool messages -#}
273
+ {%- set ns_tool_scan = namespace(stopped=false) -%}
274
+ {%- for k in range(loop.index0 + 1, loop_messages | length) -%}
275
+ {%- if ns_tool_scan.stopped -%}
276
+ {%- elif loop_messages[k]['role'] != 'tool' -%}
277
+ {%- set ns_tool_scan.stopped = true -%}
278
+ {%- else -%}
279
+ {%- set follow = loop_messages[k] -%}
280
+ {#- Resolve tool_call_id to function name -#}
281
+ {%- set ns_tname = namespace(name=follow.get('name') | default('unknown')) -%}
282
+ {%- for tc in message['tool_calls'] -%}
283
+ {%- if tc.get('id') == follow.get('tool_call_id') -%}
284
+ {%- set ns_tname.name = tc['function']['name'] -%}
285
+ {%- endif -%}
286
+ {%- endfor -%}
287
+ {#- Handle content as string or content-parts array -#}
288
+ {%- set tool_body = follow.get('content') -%}
289
+ {%- if tool_body is string -%}
290
+ {{- format_tool_response_block(ns_tname.name, tool_body) -}}
291
+ {%- elif tool_body is sequence and tool_body is not string -%}
292
+ {%- set ns_txt = namespace(s='') -%}
293
+ {%- for part in tool_body -%}
294
+ {%- if part.get('type') == 'text' -%}
295
+ {%- set ns_txt.s = ns_txt.s + (part.get('text') | default('')) -%}
296
+ {%- endif -%}
297
+ {%- endfor -%}
298
+ {{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}
299
+ {%- else -%}
300
+ {{- format_tool_response_block(ns_tname.name, tool_body) -}}
301
+ {%- endif -%}
302
+ {%- set ns_tr_out.flag = true -%}
303
+ {%- set ns.prev_message_type = 'tool_response' -%}
304
+ {%- endif -%}
305
+ {%- endfor -%}
306
+ {%- endif -%}
307
+
308
+ {%- if message['content'] is string -%}
309
+ {%- if role == 'model' -%}
310
+ {{- strip_thinking(message['content']) -}}
311
+ {%- else -%}
312
+ {{- message['content'] | trim -}}
313
+ {%- endif -%}
314
+ {%- elif message['content'] is sequence -%}
315
+ {%- for item in message['content'] -%}
316
+ {%- if item['type'] == 'text' -%}
317
+ {%- if role == 'model' -%}
318
+ {{- strip_thinking(item['text']) -}}
319
+ {%- else -%}
320
+ {{- item['text'] | trim -}}
321
+ {%- endif -%}
322
+ {%- elif item['type'] == 'image' -%}
323
+ {{- '<|image|>' -}}
324
+ {%- set ns.prev_message_type = 'image' -%}
325
+ {%- elif item['type'] == 'audio' -%}
326
+ {{- '<|audio|>' -}}
327
+ {%- set ns.prev_message_type = 'audio' -%}
328
+ {%- elif item['type'] == 'video' -%}
329
+ {{- '<|video|>' -}}
330
+ {%- set ns.prev_message_type = 'video' -%}
331
+ {%- endif -%}
332
+ {%- endfor -%}
333
+ {%- endif -%}
334
+
335
+ {%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}
336
+ {{- '<|tool_response>' -}}
337
+ {%- elif not (ns_tr_out.flag and not message.get('content')) -%}
338
+ {{- '<turn|>\n' -}}
339
+ {%- endif -%}
340
+ {%- endif -%}
341
+ {%- endfor -%}
342
+
343
+ {%- if add_generation_prompt -%}
344
+ {%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}
345
+ {{- '<|turn>model\n' -}}
346
+ {%- endif -%}
347
+ {%- endif -%}
merged_model/config.json ADDED
@@ -0,0 +1,197 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma4ForConditionalGeneration"
4
+ ],
5
+ "audio_config": {
6
+ "_name_or_path": "",
7
+ "architectures": null,
8
+ "attention_chunk_size": 12,
9
+ "attention_context_left": 13,
10
+ "attention_context_right": 0,
11
+ "attention_invalid_logits_value": -1000000000.0,
12
+ "attention_logit_cap": 50.0,
13
+ "chunk_size_feed_forward": 0,
14
+ "conv_kernel_size": 5,
15
+ "dtype": "bfloat16",
16
+ "gradient_clipping": 10000000000.0,
17
+ "hidden_act": "silu",
18
+ "hidden_size": 1024,
19
+ "id2label": {
20
+ "0": "LABEL_0",
21
+ "1": "LABEL_1"
22
+ },
23
+ "initializer_range": 0.02,
24
+ "is_encoder_decoder": false,
25
+ "label2id": {
26
+ "LABEL_0": 0,
27
+ "LABEL_1": 1
28
+ },
29
+ "model_type": "gemma4_audio",
30
+ "num_attention_heads": 8,
31
+ "num_hidden_layers": 12,
32
+ "output_attentions": false,
33
+ "output_hidden_states": false,
34
+ "output_proj_dims": 1536,
35
+ "problem_type": null,
36
+ "residual_weight": 0.5,
37
+ "return_dict": true,
38
+ "rms_norm_eps": 1e-06,
39
+ "subsampling_conv_channels": [
40
+ 128,
41
+ 32
42
+ ],
43
+ "use_clipped_linears": true
44
+ },
45
+ "audio_token_id": 258881,
46
+ "boa_token_id": 256000,
47
+ "boi_token_id": 255999,
48
+ "dtype": "bfloat16",
49
+ "eoa_token_id": 258883,
50
+ "eoa_token_index": 258883,
51
+ "eoi_token_id": 258882,
52
+ "eos_token_id": [
53
+ 1,
54
+ 106
55
+ ],
56
+ "image_token_id": 258880,
57
+ "initializer_range": 0.02,
58
+ "model_type": "gemma4",
59
+ "text_config": {
60
+ "attention_bias": false,
61
+ "attention_dropout": 0.0,
62
+ "attention_k_eq_v": false,
63
+ "bos_token_id": 2,
64
+ "dtype": "bfloat16",
65
+ "enable_moe_block": false,
66
+ "eos_token_id": 1,
67
+ "expert_intermediate_size": null,
68
+ "final_logit_softcapping": 30.0,
69
+ "global_head_dim": 512,
70
+ "head_dim": 256,
71
+ "hidden_activation": "gelu_pytorch_tanh",
72
+ "hidden_size": 2560,
73
+ "hidden_size_per_layer_input": 256,
74
+ "initializer_range": 0.02,
75
+ "intermediate_size": 10240,
76
+ "layer_types": [
77
+ "sliding_attention",
78
+ "sliding_attention",
79
+ "sliding_attention",
80
+ "sliding_attention",
81
+ "sliding_attention",
82
+ "full_attention",
83
+ "sliding_attention",
84
+ "sliding_attention",
85
+ "sliding_attention",
86
+ "sliding_attention",
87
+ "sliding_attention",
88
+ "full_attention",
89
+ "sliding_attention",
90
+ "sliding_attention",
91
+ "sliding_attention",
92
+ "sliding_attention",
93
+ "sliding_attention",
94
+ "full_attention",
95
+ "sliding_attention",
96
+ "sliding_attention",
97
+ "sliding_attention",
98
+ "sliding_attention",
99
+ "sliding_attention",
100
+ "full_attention",
101
+ "sliding_attention",
102
+ "sliding_attention",
103
+ "sliding_attention",
104
+ "sliding_attention",
105
+ "sliding_attention",
106
+ "full_attention",
107
+ "sliding_attention",
108
+ "sliding_attention",
109
+ "sliding_attention",
110
+ "sliding_attention",
111
+ "sliding_attention",
112
+ "full_attention",
113
+ "sliding_attention",
114
+ "sliding_attention",
115
+ "sliding_attention",
116
+ "sliding_attention",
117
+ "sliding_attention",
118
+ "full_attention"
119
+ ],
120
+ "max_position_embeddings": 131072,
121
+ "model_type": "gemma4_text",
122
+ "num_attention_heads": 8,
123
+ "num_experts": null,
124
+ "num_global_key_value_heads": null,
125
+ "num_hidden_layers": 42,
126
+ "num_key_value_heads": 2,
127
+ "num_kv_shared_layers": 18,
128
+ "pad_token_id": 0,
129
+ "rms_norm_eps": 1e-06,
130
+ "rope_parameters": {
131
+ "full_attention": {
132
+ "partial_rotary_factor": 0.25,
133
+ "rope_theta": 1000000.0,
134
+ "rope_type": "proportional"
135
+ },
136
+ "sliding_attention": {
137
+ "rope_theta": 10000.0,
138
+ "rope_type": "default"
139
+ }
140
+ },
141
+ "sliding_window": 512,
142
+ "tie_word_embeddings": true,
143
+ "top_k_experts": null,
144
+ "use_bidirectional_attention": null,
145
+ "use_cache": true,
146
+ "use_double_wide_mlp": false,
147
+ "vocab_size": 262144,
148
+ "vocab_size_per_layer_input": 262144
149
+ },
150
+ "tie_word_embeddings": true,
151
+ "transformers_version": "5.5.0.dev0",
152
+ "video_token_id": 258884,
153
+ "vision_config": {
154
+ "_name_or_path": "",
155
+ "architectures": null,
156
+ "attention_bias": false,
157
+ "attention_dropout": 0.0,
158
+ "chunk_size_feed_forward": 0,
159
+ "default_output_length": 280,
160
+ "dtype": "bfloat16",
161
+ "global_head_dim": 64,
162
+ "head_dim": 64,
163
+ "hidden_activation": "gelu_pytorch_tanh",
164
+ "hidden_size": 768,
165
+ "id2label": {
166
+ "0": "LABEL_0",
167
+ "1": "LABEL_1"
168
+ },
169
+ "initializer_range": 0.02,
170
+ "intermediate_size": 3072,
171
+ "is_encoder_decoder": false,
172
+ "label2id": {
173
+ "LABEL_0": 0,
174
+ "LABEL_1": 1
175
+ },
176
+ "max_position_embeddings": 131072,
177
+ "model_type": "gemma4_vision",
178
+ "num_attention_heads": 12,
179
+ "num_hidden_layers": 16,
180
+ "num_key_value_heads": 12,
181
+ "output_attentions": false,
182
+ "output_hidden_states": false,
183
+ "patch_size": 16,
184
+ "pooling_kernel_size": 3,
185
+ "position_embedding_size": 10240,
186
+ "problem_type": null,
187
+ "return_dict": true,
188
+ "rms_norm_eps": 1e-06,
189
+ "rope_parameters": {
190
+ "rope_theta": 100.0,
191
+ "rope_type": "default"
192
+ },
193
+ "standardize": false,
194
+ "use_clipped_linears": true
195
+ },
196
+ "vision_soft_tokens_per_image": 280
197
+ }
merged_model/generation_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "do_sample": true,
4
+ "eos_token_id": [
5
+ 1,
6
+ 106,
7
+ 50
8
+ ],
9
+ "pad_token_id": 0,
10
+ "temperature": 1.0,
11
+ "top_k": 64,
12
+ "top_p": 0.95,
13
+ "transformers_version": "5.5.0.dev0"
14
+ }
merged_model/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cfbd3d2f1cd71bd471c37fe2bf8546d5028d41e5736f64e1ca6c6b8893125503
3
+ size 15992595884
merged_model/processor_config.json ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "audio_ms_per_token": 40,
3
+ "audio_seq_length": 750,
4
+ "feature_extractor": {
5
+ "dither": 0.0,
6
+ "feature_extractor_type": "Gemma4AudioFeatureExtractor",
7
+ "feature_size": 128,
8
+ "fft_length": 512,
9
+ "fft_overdrive": false,
10
+ "frame_length": 320,
11
+ "hop_length": 160,
12
+ "input_scale_factor": 1.0,
13
+ "max_frequency": 8000.0,
14
+ "mel_floor": 0.001,
15
+ "min_frequency": 0.0,
16
+ "padding_side": "right",
17
+ "padding_value": 0.0,
18
+ "per_bin_mean": null,
19
+ "per_bin_stddev": null,
20
+ "preemphasis": 0.0,
21
+ "preemphasis_htk_flavor": true,
22
+ "return_attention_mask": true,
23
+ "sampling_rate": 16000
24
+ },
25
+ "image_processor": {
26
+ "do_convert_rgb": true,
27
+ "do_normalize": false,
28
+ "do_rescale": true,
29
+ "do_resize": true,
30
+ "image_mean": [
31
+ 0.0,
32
+ 0.0,
33
+ 0.0
34
+ ],
35
+ "image_processor_type": "Gemma4ImageProcessor",
36
+ "image_seq_length": 280,
37
+ "image_std": [
38
+ 1.0,
39
+ 1.0,
40
+ 1.0
41
+ ],
42
+ "max_soft_tokens": 280,
43
+ "patch_size": 16,
44
+ "pooling_kernel_size": 3,
45
+ "resample": 3,
46
+ "rescale_factor": 0.00392156862745098
47
+ },
48
+ "image_seq_length": 280,
49
+ "processor_class": "Gemma4Processor",
50
+ "video_processor": {
51
+ "do_convert_rgb": true,
52
+ "do_normalize": true,
53
+ "do_rescale": true,
54
+ "do_resize": true,
55
+ "do_sample_frames": true,
56
+ "image_mean": [
57
+ 0.0,
58
+ 0.0,
59
+ 0.0
60
+ ],
61
+ "image_std": [
62
+ 1.0,
63
+ 1.0,
64
+ 1.0
65
+ ],
66
+ "max_soft_tokens": 70,
67
+ "num_frames": 32,
68
+ "patch_size": 16,
69
+ "pooling_kernel_size": 3,
70
+ "resample": 3,
71
+ "rescale_factor": 0.00392156862745098,
72
+ "return_metadata": false,
73
+ "video_processor_type": "Gemma4VideoProcessor"
74
+ }
75
+ }
merged_model/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f
3
+ size 32169626
merged_model/tokenizer_config.json ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "audio_token": "<|audio|>",
3
+ "backend": "tokenizers",
4
+ "boa_token": "<|audio>",
5
+ "boi_token": "<|image>",
6
+ "bos_token": "<bos>",
7
+ "eoa_token": "<audio|>",
8
+ "eoc_token": "<channel|>",
9
+ "eoi_token": "<image|>",
10
+ "eos_token": "<eos>",
11
+ "eot_token": "<turn|>",
12
+ "escape_token": "<|\"|>",
13
+ "etc_token": "<tool_call|>",
14
+ "etd_token": "<tool|>",
15
+ "etr_token": "<tool_response|>",
16
+ "extra_special_tokens": [
17
+ "<|video|>"
18
+ ],
19
+ "image_token": "<|image|>",
20
+ "mask_token": "<mask>",
21
+ "model_max_length": 1000000000000000019884624838656,
22
+ "pad_token": "<pad>",
23
+ "padding_side": "left",
24
+ "processor_class": "Gemma4Processor",
25
+ "response_schema": {
26
+ "type": "object",
27
+ "properties": {
28
+ "role": {
29
+ "const": "assistant"
30
+ },
31
+ "thinking": {
32
+ "type": "string"
33
+ },
34
+ "content": {
35
+ "type": "string"
36
+ },
37
+ "tool_calls": {
38
+ "x-regex-iterator": "<\\|tool_call>(.*?)<tool_call\\|>",
39
+ "type": "array",
40
+ "items": {
41
+ "type": "object",
42
+ "properties": {
43
+ "type": {
44
+ "const": "function"
45
+ },
46
+ "function": {
47
+ "type": "object",
48
+ "x-regex": "call\\:(?P<name>\\w+)(?P<arguments>\\{.*\\})",
49
+ "properties": {
50
+ "name": {
51
+ "type": "string"
52
+ },
53
+ "arguments": {
54
+ "type": "object",
55
+ "x-parser": "gemma4-tool-call",
56
+ "additionalProperties": {}
57
+ }
58
+ }
59
+ }
60
+ }
61
+ }
62
+ }
63
+ },
64
+ "x-regex": "(\\<\\|channel\\>thought\\n(?P<thinking>.*?)\\<channel\\|\\>)?(?P<tool_calls>\\<\\|tool_call\\>.*\\<tool_call\\|\\>)?(?P<content>(?:(?!\\<turn\\|\\>)(?!\\<\\|tool_response\\>).)+)?(?:\\<turn\\|\\>|\\<\\|tool_response\\>)?"
65
+ },
66
+ "soc_token": "<|channel>",
67
+ "sot_token": "<|turn>",
68
+ "stc_token": "<|tool_call>",
69
+ "std_token": "<|tool>",
70
+ "str_token": "<|tool_response>",
71
+ "think_token": "<|think|>",
72
+ "tokenizer_class": "GemmaTokenizer",
73
+ "unk_token": "<unk>"
74
+ }
ollama/Modelfile ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ FROM ../merged_model
2
+ TEMPLATE """{{ if .System }}<|im_start|>system
3
+ {{ .System }}<|im_end|>
4
+ {{ end }}{{ if .Prompt }}<|im_start|>user
5
+ {{ .Prompt }}<|im_end|>
6
+ {{ end }}<|im_start|>assistant
7
+ """
8
+ PARAMETER temperature 0.1
9
+ PARAMETER num_ctx 8192
10
+ PARAMETER stop "<|im_end|>"
11
+ PARAMETER stop "<|endoftext|>"
ollama/create_ollama_model.ps1 ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ param(
2
+ [string]$ModelName = "lumynax-infused-gemma-e4b"
3
+ )
4
+
5
+ $ErrorActionPreference = "Stop"
6
+ Set-StrictMode -Version Latest
7
+
8
+ $scriptDir = Split-Path -Parent $MyInvocation.MyCommand.Path
9
+ $modelfilePath = Join-Path $scriptDir "Modelfile"
10
+
11
+ if (-not (Get-Command ollama -ErrorAction SilentlyContinue)) {
12
+ throw "The `ollama` CLI is not installed. Install Ollama first."
13
+ }
14
+
15
+ & ollama create $ModelName -f $modelfilePath
16
+ if ($LASTEXITCODE -ne 0) {
17
+ exit $LASTEXITCODE
18
+ }
19
+
20
+ Write-Output "Created Ollama model: $ModelName"
21
+ Write-Output "Run it with: ollama run $ModelName"
quickstart.py ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+
3
+ import argparse
4
+ import json
5
+ from pathlib import Path
6
+
7
+ import torch
8
+ from transformers import AutoModelForMultimodalLM, AutoProcessor
9
+
10
+ MODEL_TITLE = "LumynaX Infused Gemma E4B Model"
11
+ SUPPORTED_MODALITIES = ('text', 'image', 'audio')
12
+ DEFAULT_ENABLE_THINKING = True
13
+ DEFAULT_SYSTEM_PROMPT = 'You are LumynaX operating from the LumynaX Infused Gemma E4B Model package identity. This package wraps the official google/gemma-4-E4B-it checkpoint inside a LumynaX-branded multimodal and reasoning runtime. Always identify yourself as LumynaX when asked who you are. Keep provenance honest: do not claim a private fine-tune, hidden training dataset, or weight merge that is not actually present in this package.'
14
+
15
+
16
+ def _build_parser() -> argparse.ArgumentParser:
17
+ parser = argparse.ArgumentParser(
18
+ description=(
19
+ f"Run a local Gemma E4B quickstart for {MODEL_TITLE}. "
20
+ f"Supported modalities: text, image, audio."
21
+ )
22
+ )
23
+ parser.add_argument("--mode", choices=["text", "image", "audio"], default="text")
24
+ parser.add_argument(
25
+ "--prompt",
26
+ default="Explain in two short bullet points what this local package is.",
27
+ help="Text instruction to send to the model.",
28
+ )
29
+ parser.add_argument("--image", default="", help="Local image path or image URL for --mode image.")
30
+ parser.add_argument("--audio", default="", help="Local audio path or audio URL for --mode audio.")
31
+ parser.add_argument("--max-new-tokens", type=int, default=256)
32
+ parser.add_argument(
33
+ "--thinking",
34
+ action=argparse.BooleanOptionalAction,
35
+ default=DEFAULT_ENABLE_THINKING,
36
+ help="Enable Gemma reasoning mode.",
37
+ )
38
+ return parser
39
+
40
+
41
+ def _message_content(args: argparse.Namespace) -> list[dict[str, str]]:
42
+ if args.mode == "text":
43
+ return [{"type": "text", "text": args.prompt}]
44
+ if args.mode == "image":
45
+ if not args.image:
46
+ raise SystemExit("--image is required when --mode image is used.")
47
+ image_ref = args.image.strip()
48
+ return [
49
+ {"type": "image", "url": image_ref},
50
+ {"type": "text", "text": args.prompt},
51
+ ]
52
+ if not args.audio:
53
+ raise SystemExit("--audio is required when --mode audio is used.")
54
+ audio_ref = args.audio.strip()
55
+ return [
56
+ {"type": "audio", "audio": audio_ref},
57
+ {"type": "text", "text": args.prompt},
58
+ ]
59
+
60
+
61
+ def main() -> None:
62
+ args = _build_parser().parse_args()
63
+ model_dir = Path(__file__).resolve().parent / "merged_model"
64
+ if not model_dir.exists():
65
+ raise SystemExit(f"Expected merged_model/ at {model_dir}")
66
+
67
+ processor = AutoProcessor.from_pretrained(model_dir)
68
+ model = AutoModelForMultimodalLM.from_pretrained(
69
+ model_dir,
70
+ dtype="auto",
71
+ device_map="auto",
72
+ )
73
+ messages = [
74
+ {
75
+ "role": "system",
76
+ "content": [
77
+ {
78
+ "type": "text",
79
+ "text": DEFAULT_SYSTEM_PROMPT,
80
+ }
81
+ ],
82
+ },
83
+ {
84
+ "role": "user",
85
+ "content": _message_content(args),
86
+ },
87
+ ]
88
+ inputs = processor.apply_chat_template(
89
+ messages,
90
+ tokenize=True,
91
+ return_dict=True,
92
+ return_tensors="pt",
93
+ add_generation_prompt=True,
94
+ enable_thinking=args.thinking,
95
+ ).to(model.device)
96
+ input_len = inputs["input_ids"].shape[-1]
97
+
98
+ with torch.inference_mode():
99
+ outputs = model.generate(**inputs, max_new_tokens=args.max_new_tokens)
100
+
101
+ response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
102
+ parsed = processor.parse_response(response) if hasattr(processor, "parse_response") else response
103
+ if isinstance(parsed, str):
104
+ print(parsed)
105
+ return
106
+ print(json.dumps(parsed, indent=2, ensure_ascii=False))
107
+
108
+
109
+ if __name__ == "__main__":
110
+ main()
release_export_manifest.json ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "checksums": "checksums.sha256",
4
+ "gguf": null,
5
+ "hf_space_app": "hf_space/app.py",
6
+ "hf_space_dir": "hf_space",
7
+ "hf_space_readme": "hf_space/README.md",
8
+ "hf_space_requirements": "hf_space/requirements.txt",
9
+ "license": "LICENSE.txt",
10
+ "lumynax_package_identity": "merged_model/LUMYNAX_PACKAGE_IDENTITY.txt",
11
+ "merged_model": "merged_model",
12
+ "ollama_create_script": "ollama/create_ollama_model.ps1",
13
+ "ollama_modelfile": "ollama/Modelfile",
14
+ "quantized_gguf": null,
15
+ "quickstart": "quickstart.py",
16
+ "readme": "README.md",
17
+ "requirements": "requirements.txt",
18
+ "training_summary": "artifacts/release_training_summary.json",
19
+ "version": "VERSION.txt"
20
+ },
21
+ "capabilities": {
22
+ "reasoning_enabled": true,
23
+ "supported_modalities": [
24
+ "text",
25
+ "image",
26
+ "audio"
27
+ ]
28
+ },
29
+ "delivery": "standalone_hf_base_release",
30
+ "distribution": {
31
+ "hf_space": {
32
+ "app": "hf_space/app.py",
33
+ "default_model_repo_id": "AbteeXAILab/lumynax-infused-gemma-e4b",
34
+ "directory": "hf_space",
35
+ "model_repo_env_var": "LUMYNAX_MODEL_REPO_ID",
36
+ "readme": "hf_space/README.md",
37
+ "requirements": "hf_space/requirements.txt",
38
+ "status": "private_reasoning_multimodal_demo"
39
+ },
40
+ "ollama": {
41
+ "create_script": "ollama/create_ollama_model.ps1",
42
+ "modelfile": "ollama/Modelfile",
43
+ "recommended_model_name": "lumynax-infused-gemma-e4b",
44
+ "status": "not_validated_for_multimodal_base_release"
45
+ }
46
+ },
47
+ "family": null,
48
+ "generated_at": "2026-04-11T23:33:11.251732+00:00",
49
+ "lumynax_package_identity": {
50
+ "chat_template": "merged_model/chat_template.jinja",
51
+ "hardcoded": true,
52
+ "identity_file": "merged_model/LUMYNAX_PACKAGE_IDENTITY.txt",
53
+ "runtime_wrappers": [
54
+ "quickstart.py",
55
+ "hf_space/app.py"
56
+ ]
57
+ },
58
+ "manifest_version": 2,
59
+ "model_title": "LumynaX Infused Gemma E4B Model",
60
+ "package_state": "base_weights_hydrated",
61
+ "public_identity": {
62
+ "model_name": "LumynaX",
63
+ "organization": "AbteeX AI Labs",
64
+ "region": "Aotearoa New Zealand"
65
+ },
66
+ "release_version": "v1",
67
+ "runtime": {
68
+ "delivery_mode": "standalone_base_release",
69
+ "preferred_backend": "transformers_multimodal",
70
+ "prompt_format": "huggingface_chat_template",
71
+ "quickstart_command": "python quickstart.py"
72
+ },
73
+ "upstream_model": {
74
+ "kind": "official_base_weights",
75
+ "lumynax_weight_adaptation_applied": false,
76
+ "provider": "Hugging Face",
77
+ "repo_id": "google/gemma-4-E4B-it"
78
+ }
79
+ }
requirements.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ accelerate>=1.13
2
+ huggingface-hub>=1.8
3
+ numba>=0.65
4
+ pillow>=10.0
5
+ safetensors>=0.6
6
+ torch>=2.9
7
+ torchvision>=0.24
8
+ transformers>=5.5.3
9
+ librosa>=0.11