[model] support encoder-free unified Text/Vision/Audio model (#27167)

2026-06-03 23:58:06 +08:00
parent f65aae8493
commit fa5c8a3101
11 changed files with 816 additions and 5 deletions
@@ -47,6 +47,11 @@ Gemma 4 is Google's next-generation family of open models, building on the Gemma
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Dense</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>~4B</td>
    </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>[google/gemma-4-12B-it](https://huggingface.co/google/gemma-4-12B-it)</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Dense</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>12B</td>
+    </tr>
    <tr>
      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>[google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it)</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Dense</td>
@@ -70,6 +75,8 @@ pip install 'git+https://github.com/sgl-project/sglang.git#subdirectory=python'

 # Install transformers with Gemma 4 support
 pip install 'git+https://github.com/huggingface/transformers.git@91b1ab1fdfa81a552644a92fbe3e8d88de40e167'
+# gemma-4-12B-it (unified) additionally requires a transformers build that
+# includes the gemma4_unified model family (>= 5.10).

 # Or use Docker AMD64
 docker pull lmsysorg/sglang:gemma4 # CUDA 12.9
@@ -93,6 +100,7 @@ For the full Docker setup and other installation methods, please refer to the [o
 ### 3.2 Configuration Tips

 - SGLang automatically selects the Triton attention backend for Gemma 4 models (required for bidirectional image-token attention during prefill).
+- **Attention backend on Blackwell (B200/sm100)**: SGLang defaults to the `trtllm_mha` backend on sm100, which is fastest for text but applies *causal* attention to image tokens. For multimodal (image) workloads on B200, pass `--attention-backend triton` to restore bidirectional image-token attention and full vision quality. Text-only and audio workloads are unaffected by the default.
 - For the 26B-A4B MoE model, consider `--tp 2` for high-throughput workloads.
 - **Speculative Decoding (MTP)**: Each Gemma 4 variant ships with a paired `*-assistant` draft model that enables NEXTN multi-token prediction. Enable it via the selector above, or pass `--speculative-algorithm NEXTN --speculative-draft-model-path google/gemma-4-<variant>-it-assistant --speculative-num-steps 5 --speculative-num-draft-tokens 6 --speculative-eagle-topk 1`. MTP can significantly reduce latency for interactive use cases. The 26B-A4B MoE model requires `--tp 2` when MTP is enabled.
 - Hardware requirements:
@@ -121,6 +129,11 @@ For the full Docker setup and other installation methods, please refer to the [o
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>1x H200 / 1x MI300X / 1x MI325X / 1x MI355X</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>1</td>
    </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-12B-it</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>1x H200 / 1x B200</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>1</td>
+    </tr>
    <tr>
      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-31B-it</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>2x H200 / 1x MI300X / 1x MI325X / 1x MI355X</td>
@@ -188,6 +201,18 @@ sglang serve \
  --mem-fraction-static 0.85
 ```

+```bash Command
+# Gemma 4 12B + MTP (~35% faster single-stream decode on H200)
+sglang serve \
+  --model-path google/gemma-4-12B-it \
+  --speculative-algorithm NEXTN \
+  --speculative-draft-model-path google/gemma-4-12B-it-assistant \
+  --speculative-num-steps 5 \
+  --speculative-num-draft-tokens 6 \
+  --speculative-eagle-topk 1 \
+  --mem-fraction-static 0.85
+```
+
 ```bash Command
 # Gemma 4 31B + MTP
 sglang serve \
@@ -498,6 +523,58 @@ Tool Call: get_weather

 </details>

+### 4.5 Audio Input
+
+The audio-capable Gemma 4 variants (`gemma-4-E2B-it`, `gemma-4-E4B-it`, `gemma-4-12B-it`) accept raw audio alongside text. Pass the waveform as a base64 `audio_url` data URI (16 kHz mono WAV works well):
+
+```python Example
+import base64
+from openai import OpenAI
+
+client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
+
+with open("sample.wav", "rb") as f:
+    audio_b64 = base64.b64encode(f.read()).decode()
+
+response = client.chat.completions.create(
+    model="google/gemma-4-12B-it",
+    messages=[
+        {
+            "role": "user",
+            "content": [
+                {"type": "audio_url", "audio_url": {"url": f"data:audio/wav;base64,{audio_b64}"}},
+                {"type": "text", "text": "Transcribe the speech in this audio exactly."},
+            ],
+        }
+    ],
+    max_tokens=256,
+    temperature=0,
+)
+
+print(response.choices[0].message.content)
+```
+
+<details>
+<summary>Example Output</summary>
+
+```text Output
+Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.
+```
+
+</details>
+
+For best ASR quality, use the recommended transcription prompt structure:
+
+```text Prompt
+Transcribe the following speech segment in {LANGUAGE} into {LANGUAGE} text.
+
+Follow these specific instructions for formatting the answer:
+* Only output the transcription, with no newlines.
+* When transcribing numbers, write the digits, i.e. write 1.7 and not one point seven, and write 3 instead of three.
+```
+
+For speech translation (AST), ask for the transcription in the source language first, then the translation: *"Transcribe the following speech segment in {SOURCE_LANGUAGE}, then translate it into {TARGET_LANGUAGE}. ..."*
+
 ## 5. Benchmark

 ### 5.1 Speed Benchmark
@@ -1033,6 +1110,201 @@ Median ITL (ms):                         29.31
 ==================================================
 ```

+#### gemma-4-12B-it (1x H200, TP=1)
+
+Server Launch Command:
+```bash Command
+sglang serve --model-path google/gemma-4-12B-it
+```
+
+**Latency Benchmark (Text)**
+
+```bash Command
+python3 -m sglang.bench_serving --backend sglang \
+  --host 0.0.0.0 --port 30000 \
+  --dataset-name random --num-prompts 10 --max-concurrency 1
+```
+
+```text Output
+============ Serving Benchmark Result ============
+Backend:                                 sglang
+Max request concurrency:                 1
+Successful requests:                     10
+Benchmark duration (s):                  38.66
+Total input tokens:                      6101
+Total generated tokens:                  4220
+Request throughput (req/s):              0.26
+Output token throughput (tok/s):         109.15
+Total token throughput (tok/s):          266.94
+Mean TTFT (ms):                          33.08
+Median TTFT (ms):                        33.71
+Mean TPOT (ms):                          9.02
+Median ITL (ms):                         9.19
+==================================================
+```
+
+**Latency Benchmark (Image)**
+
+```bash Command
+python3 -m sglang.bench_serving --backend sglang-oai-chat \
+  --host 0.0.0.0 --port 30000 \
+  --dataset-name image --image-count 2 --image-resolution 720p \
+  --random-input-len 128 --random-output-len 1024 \
+  --num-prompts 10 --max-concurrency 1
+```
+
+```text Output
+============ Serving Benchmark Result ============
+Backend:                                 sglang-oai-chat
+Max request concurrency:                 1
+Successful requests:                     10
+Benchmark duration (s):                  39.36
+Total input vision tokens:               5320
+Total generated tokens:                  4220
+Request throughput (req/s):              0.25
+Output token throughput (tok/s):         107.23
+Total token throughput (tok/s):          263.62
+Mean TTFT (ms):                          94.98
+Median TTFT (ms):                        97.33
+Mean TPOT (ms):                          9.08
+Median ITL (ms):                         9.17
+==================================================
+```
+
+**Throughput Benchmark (Text)**
+
+```bash Command
+python3 -m sglang.bench_serving --backend sglang \
+  --host 0.0.0.0 --port 30000 \
+  --dataset-name random --num-prompts 1000 --max-concurrency 100
+```
+
+```text Output
+============ Serving Benchmark Result ============
+Backend:                                 sglang
+Max request concurrency:                 100
+Successful requests:                     1000
+Benchmark duration (s):                  130.44
+Total input tokens:                      512842
+Total generated tokens:                  510855
+Request throughput (req/s):              7.67
+Output token throughput (tok/s):         3916.46
+Total token throughput (tok/s):          7848.15
+Mean TTFT (ms):                          207.49
+Median TTFT (ms):                        76.95
+Mean TPOT (ms):                          24.38
+Median ITL (ms):                         17.89
+==================================================
+```
+
+**Throughput Benchmark (Image)**
+
+```text Output
+============ Serving Benchmark Result ============
+Backend:                                 sglang-oai-chat
+Max request concurrency:                 100
+Successful requests:                     1000
+Benchmark duration (s):                  147.57
+Total input tokens:                      619609
+Total input vision tokens:               532000
+Total generated tokens:                  510855
+Request throughput (req/s):              6.78
+Output token throughput (tok/s):         3461.79
+Total token throughput (tok/s):          7660.54
+Mean TTFT (ms):                          438.40
+Median TTFT (ms):                        129.83
+Mean TPOT (ms):                          27.12
+Median ITL (ms):                         19.16
+==================================================
+```
+
+#### gemma-4-12B-it (1x B200, TP=1)
+
+Server Launch Command:
+```bash Command
+# Text/audio: the sm100 default (trtllm_mha) is fastest.
+# For image workloads add --attention-backend triton (bidirectional image attention).
+sglang serve --model-path google/gemma-4-12B-it --attention-backend triton
+```
+
+**Latency Benchmark (Text)**
+
+```text Output
+============ Serving Benchmark Result ============
+Backend:                                 sglang
+Max request concurrency:                 1
+Successful requests:                     10
+Benchmark duration (s):                  30.46
+Output token throughput (tok/s):         138.55
+Total token throughput (tok/s):          338.85
+Mean TTFT (ms):                          28.14
+Median TTFT (ms):                        29.74
+Mean TPOT (ms):                          7.08
+Median ITL (ms):                         7.26
+==================================================
+```
+
+**Latency Benchmark (Image)**
+
+```text Output
+============ Serving Benchmark Result ============
+Backend:                                 sglang-oai-chat
+Max request concurrency:                 1
+Successful requests:                     10
+Benchmark duration (s):                  31.43
+Total input vision tokens:               5320
+Total generated tokens:                  4220
+Request throughput (req/s):              0.32
+Output token throughput (tok/s):         134.26
+Total token throughput (tok/s):          329.57
+Mean TTFT (ms):                          115.51
+Median TTFT (ms):                        74.27
+Mean TPOT (ms):                          7.14
+Median ITL (ms):                         7.24
+==================================================
+```
+
+**Throughput Benchmark (Text)**
+
+```text Output
+============ Serving Benchmark Result ============
+Backend:                                 sglang
+Max request concurrency:                 100
+Successful requests:                     1000
+Benchmark duration (s):                  92.94
+Request throughput (req/s):              10.76
+Output token throughput (tok/s):         5496.55
+Total token throughput (tok/s):          11014.49
+Mean TTFT (ms):                          120.89
+Median TTFT (ms):                        45.00
+Mean TPOT (ms):                          17.23
+Median ITL (ms):                         14.30
+==================================================
+```
+
+**Throughput Benchmark (Image)**
+
+```text Output
+============ Serving Benchmark Result ============
+Backend:                                 sglang-oai-chat
+Max request concurrency:                 100
+Successful requests:                     998
+Benchmark duration (s):                  107.82
+Total input tokens:                      617971
+Total input vision tokens:               530936
+Total generated tokens:                  508951
+Request throughput (req/s):              9.26
+Output token throughput (tok/s):         4720.29
+Total token throughput (tok/s):          10451.68
+Mean TTFT (ms):                          425.89
+Median TTFT (ms):                        109.57
+Mean TPOT (ms):                          19.45
+Median ITL (ms):                         15.11
+==================================================
+```
+
+> **Performance tuning**: On B200, raising `--scheduler-recv-interval` to 16 lifted text throughput from 5497 to 5673 tok/s output (≈ +3%) at concurrency 100 with no accuracy change, by reducing the scheduler's per-step Python overhead. It is a safe, low-risk knob for high-concurrency serving.
+
 ### 5.2 Accuracy Benchmark

 **Test Environment:**
@@ -1078,6 +1350,14 @@ Median ITL (ms):                         29.31
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>0.825</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>**0.810**</td>
    </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-12B-it</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>0.784</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>0.888</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>0.946</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>0.861</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>**0.859**</td>
+    </tr>
    <tr>
      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-31B-it</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>0.878</td>
@@ -1131,6 +1411,13 @@ Median ITL (ms):                         29.31
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>4.174</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>4672.030</td>
    </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-12B-it</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>0.431</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>0.052</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>55.105</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>6580.229</td>
+    </tr>
    <tr>
      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-31B-it</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>0.805</td>
@@ -1148,6 +1435,8 @@ Median ITL (ms):                         29.31
  </tbody>
 </table>

+> **Note**: These GSM8K numbers use the raw few-shot completion harness (`sglang.test.few_shot_gsm8k`). `gemma-4-12B-it` is reasoning-oriented and is under-elicited by raw few-shot prompting; with the chat template it scores **0.950** on the same 1319 GSM8K test questions (`sglang.test.run_eval --eval-name gsm8k`).
+
 #### MMMU

 <table style={{width: "100%", borderCollapse: "collapse", tableLayout: "fixed"}}>
@@ -1170,6 +1459,10 @@ Median ITL (ms):                         29.31
      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-E4B-it</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>**0.396**</td>
    </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-12B-it</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>**0.683**</td>
+    </tr>
    <tr>
      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-31B-it</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>**0.589**</td>
@@ -1196,6 +1489,12 @@ Median ITL (ms):                         29.31
 {"Overall-Art and Design": {"num": 120, "acc": 0.458}, "Art": {"num": 30, "acc": 0.433}, "Art_Theory": {"num": 30, "acc": 0.567}, "Design": {"num": 30, "acc": 0.667}, "Music": {"num": 30, "acc": 0.167}, "Overall-Business": {"num": 150, "acc": 0.287}, "Accounting": {"num": 30, "acc": 0.233}, "Economics": {"num": 30, "acc": 0.467}, "Finance": {"num": 30, "acc": 0.133}, "Manage": {"num": 30, "acc": 0.3}, "Marketing": {"num": 30, "acc": 0.3}, "Overall-Science": {"num": 150, "acc": 0.28}, "Biology": {"num": 30, "acc": 0.333}, "Chemistry": {"num": 30, "acc": 0.133}, "Geography": {"num": 30, "acc": 0.4}, "Math": {"num": 30, "acc": 0.2}, "Physics": {"num": 30, "acc": 0.333}, "Overall-Health and Medicine": {"num": 150, "acc": 0.427}, "Basic_Medical_Science": {"num": 30, "acc": 0.4}, "Clinical_Medicine": {"num": 30, "acc": 0.533}, "Diagnostics_and_Laboratory_Medicine": {"num": 30, "acc": 0.4}, "Pharmacy": {"num": 30, "acc": 0.4}, "Public_Health": {"num": 30, "acc": 0.4}, "Overall-Humanities and Social Science": {"num": 120, "acc": 0.7}, "History": {"num": 30, "acc": 0.633}, "Literature": {"num": 30, "acc": 0.867}, "Sociology": {"num": 30, "acc": 0.733}, "Psychology": {"num": 30, "acc": 0.567}, "Overall-Tech and Engineering": {"num": 210, "acc": 0.324}, "Agriculture": {"num": 30, "acc": 0.533}, "Architecture_and_Engineering": {"num": 30, "acc": 0.3}, "Computer_Science": {"num": 30, "acc": 0.367}, "Electronics": {"num": 30, "acc": 0.133}, "Energy_and_Power": {"num": 30, "acc": 0.4}, "Materials": {"num": 30, "acc": 0.2}, "Mechanical_Engineering": {"num": 30, "acc": 0.333}, "Overall": {"num": 900, "acc": 0.396}}
 ```

+**gemma-4-12B-it**
+
+```json Config
+{"Overall-Art and Design": {"num": 120, "acc": 0.667}, "Art": {"num": 30, "acc": 0.7}, "Art_Theory": {"num": 30, "acc": 0.867}, "Design": {"num": 30, "acc": 0.767}, "Music": {"num": 30, "acc": 0.333}, "Overall-Business": {"num": 150, "acc": 0.747}, "Accounting": {"num": 30, "acc": 0.767}, "Economics": {"num": 30, "acc": 0.767}, "Finance": {"num": 30, "acc": 0.633}, "Manage": {"num": 30, "acc": 0.7}, "Marketing": {"num": 30, "acc": 0.867}, "Overall-Science": {"num": 150, "acc": 0.647}, "Biology": {"num": 30, "acc": 0.6}, "Chemistry": {"num": 30, "acc": 0.633}, "Geography": {"num": 30, "acc": 0.567}, "Math": {"num": 30, "acc": 0.6}, "Physics": {"num": 30, "acc": 0.833}, "Overall-Health and Medicine": {"num": 150, "acc": 0.68}, "Basic_Medical_Science": {"num": 30, "acc": 0.667}, "Clinical_Medicine": {"num": 30, "acc": 0.633}, "Diagnostics_and_Laboratory_Medicine": {"num": 30, "acc": 0.267}, "Pharmacy": {"num": 30, "acc": 0.833}, "Public_Health": {"num": 30, "acc": 1.0}, "Overall-Humanities and Social Science": {"num": 120, "acc": 0.817}, "History": {"num": 30, "acc": 0.8}, "Literature": {"num": 30, "acc": 0.9}, "Sociology": {"num": 30, "acc": 0.8}, "Psychology": {"num": 30, "acc": 0.767}, "Overall-Tech and Engineering": {"num": 210, "acc": 0.6}, "Agriculture": {"num": 30, "acc": 0.467}, "Architecture_and_Engineering": {"num": 30, "acc": 0.667}, "Computer_Science": {"num": 30, "acc": 0.733}, "Electronics": {"num": 30, "acc": 0.567}, "Energy_and_Power": {"num": 30, "acc": 0.667}, "Materials": {"num": 30, "acc": 0.567}, "Mechanical_Engineering": {"num": 30, "acc": 0.533}, "Overall": {"num": 900, "acc": 0.683}}
+```
+
 **gemma-4-31B-it**

 ```json Config
@@ -1238,6 +1537,12 @@ Median ITL (ms):                         29.31
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>0.366</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>2.46</td>
    </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-12B-it</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Supported (see §4.5)</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>—</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>—</td>
+    </tr>
    <tr>
      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-31B-it</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Not Supported</td>
@@ -1283,6 +1588,12 @@ Median ITL (ms):                         29.31
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>0.8707s</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>16.20</td>
    </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-12B-it</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Supported (see §4.5)</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>—</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>—</td>
+    </tr>
    <tr>
      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>gemma-4-31B-it</td>
      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Not Supported</td>
@@ -6,6 +6,7 @@ export const Gemma4Deployment = () => {
      items: [
        { id: 'e2b', label: 'E2B (~2B)', default: false },
        { id: 'e4b', label: 'E4B (~4B)', default: true },
+        { id: '12b', label: '12B (Dense)', default: false },
        { id: '31b', label: '31B (Dense)', default: false },
        { id: '26b-a4b', label: '26B-A4B (MoE)', default: false },
      ]
@@ -56,12 +57,14 @@ export const Gemma4Deployment = () => {
    h200: {
      e2b: { tp: 1, mem: 0.85 },
      e4b: { tp: 1, mem: 0.85 },
+      '12b': { tp: 1, mem: 0.85 },
      '31b': { tp: 2, mem: 0.85 },
      '26b-a4b': { tp: 1, mem: 0.85 },
    },
    b200: {
      e2b: { tp: 1, mem: 0.9 },
      e4b: { tp: 1, mem: 0.9 },
+      '12b': { tp: 1, mem: 0.9 },
      '31b': { tp: 1, mem: 0.9 },
      '26b-a4b': { tp: 1, mem: 0.9 },
    },
@@ -82,6 +85,7 @@ export const Gemma4Deployment = () => {
    const modelNames = {
      'e2b': 'google/gemma-4-E2B-it',
      'e4b': 'google/gemma-4-E4B-it',
+      '12b': 'google/gemma-4-12B-it',
      '31b': 'google/gemma-4-31B-it',
      '26b-a4b': 'google/gemma-4-26B-A4B-it',
    };
@@ -25,8 +25,10 @@ def _resolve_speculative_algorithm_alias(
        cfg = get_config(
            speculative_draft_model_path, trust_remote_code=trust_remote_code, **kwargs
        )
-        is_gemma4_draft = "Gemma4AssistantForCausalLM" in (
-            getattr(cfg, "architectures", None) or []
+        draft_archs = getattr(cfg, "architectures", None) or []
+        is_gemma4_draft = any(
+            arch in ("Gemma4AssistantForCausalLM", "Gemma4UnifiedAssistantForCausalLM")
+            for arch in draft_archs
        )

    if speculative_algorithm == "EAGLE3" and is_gemma4_draft:
@@ -531,6 +531,7 @@ class ModelConfig:
            "MiMoV2MTP",
            "Gemma4ForCausalLM",
            "Gemma4ForConditionalGeneration",
+            "Gemma4UnifiedForConditionalGeneration",
        ]

    def _detect_attention_sinks(self) -> bool:
@@ -1515,6 +1516,7 @@ multimodal_model_archs = [
    "Gemma3ForConditionalGeneration",
    "Gemma3nForConditionalGeneration",
    "Gemma4ForConditionalGeneration",
+    "Gemma4UnifiedForConditionalGeneration",
    "Glm4vForConditionalGeneration",
    "Glm4vMoeForConditionalGeneration",
    "GlmOcrForConditionalGeneration",
@@ -1689,6 +1691,7 @@ def is_hybrid_swa_model(
        "Step3p7ForConditionalGeneration",
        "Gemma4ForCausalLM",
        "Gemma4ForConditionalGeneration",
+        "Gemma4UnifiedForConditionalGeneration",
        "LagunaForCausalLM",
    }
    if any(arch in hybrid_swa_archs for arch in model_architectures):
@@ -1752,6 +1755,7 @@ def get_hybrid_layer_ids(
    elif (
        "Gemma4ForCausalLM" in model_architectures
        or "Gemma4ForConditionalGeneration" in model_architectures
+        or "Gemma4UnifiedForConditionalGeneration" in model_architectures
    ):
        layer_types = getattr(hf_text_config, "layer_types", [])
        swa_attention_layer_ids = [
@@ -174,7 +174,8 @@ class OpenAIServingChat(OpenAIServingBase):
        self.is_gemma4 = (
            hasattr(self.tokenizer_manager.model_config, "hf_config")
            and hasattr(self.tokenizer_manager.model_config.hf_config, "model_type")
-            and self.tokenizer_manager.model_config.hf_config.model_type == "gemma4"
+            and self.tokenizer_manager.model_config.hf_config.model_type
+            in ("gemma4", "gemma4_unified")
        )

        # Which Python-based chat encoder (if any) bypasses apply_chat_template.
@@ -397,4 +397,8 @@ class Gemma4AssistantForCausalLM(Gemma4ForCausalLM):
        )


-EntryClass = Gemma4AssistantForCausalLM
+class Gemma4UnifiedAssistantForCausalLM(Gemma4AssistantForCausalLM):
+    """Gemma 4 unified MTP assistant; text path identical to the gemma4 assistant."""
+
+
+EntryClass = [Gemma4AssistantForCausalLM, Gemma4UnifiedAssistantForCausalLM]
@@ -0,0 +1,438 @@
+# Copyright 2026 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Gemma4 *Unified* (encoder-free) multimodal model for SGLang.
+
+The unified Gemma4 family (e.g. ``google/gemma-4-12B-it``, arch
+``Gemma4UnifiedForConditionalGeneration``, ``model_type="gemma4_unified"``)
+shares the Gemma4 *text* decoder verbatim but replaces both modality towers
+with light "encoder-free" projection pipelines:
+
+* **Vision** — raw merged pixel patches are projected directly into LM space:
+  ``LN -> Dense -> LN -> +factorized_posemb -> LN`` (``Gemma4UnifiedVisionEmbedder``)
+  followed by ``RMSNorm -> Linear`` (``Gemma4UnifiedMultimodalEmbedder``).
+  There is **no** SigLIP attention tower.
+* **Audio** — raw 16 kHz waveform is chunked into fixed ``audio_samples_per_token``
+  frames and projected straight through ``RMSNorm -> Linear``. There is **no**
+  conformer/USM encoder and no mel spectrogram.
+
+Because the text path is identical to ``gemma4``, we reuse ``Gemma4TextModel``
+and subclass ``Gemma4ForConditionalGeneration`` (reusing its ``forward``,
+bidirectional-image ``prepare_attn_masks`` and PP/embed plumbing), overriding
+only construction, per-modality feature extraction and weight loading.
+"""
+
+import logging
+import re
+from typing import Iterable, List, Optional, Set, Tuple
+
+import torch
+from torch import nn
+from transformers import PreTrainedModel
+
+from sglang.srt.distributed import get_pp_group
+from sglang.srt.layers.layernorm import Gemma4RMSNorm
+from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.utils import PPMissingLayer
+from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+from sglang.srt.managers.schedule_batch import (
+    MultimodalDataItem,
+    flatten_nested_list,
+)
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.models.gemma4_causal import Gemma4TextModel, pp_filter_load_weight
+from sglang.srt.models.gemma4_mm import Gemma4ForConditionalGeneration
+from sglang.srt.utils import add_prefix
+
+logger = logging.getLogger(__name__)
+
+
+class Gemma4UnifiedVisionEmbedder(nn.Module):
+    """Encoder-free vision embedder.
+
+    Projects raw merged pixel patches ``(..., model_patch_size**2 * 3)`` into
+    ``mm_embed_dim`` via ``LN1 -> Dense -> LN2``, adds factorized 2D positional
+    embeddings, and applies a final ``LN``.  Mirrors HF
+    ``Gemma4UnifiedVisionEmbedder``; runs on the first PP rank only, so it uses
+    plain (un-sharded) ``nn`` modules.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        patch_dim = config.model_patch_size**2 * 3  # 48*48*3 = 6912
+        mm_embed_dim = config.mm_embed_dim
+
+        self.patch_ln1 = nn.LayerNorm(patch_dim)
+        self.patch_dense = nn.Linear(patch_dim, mm_embed_dim)
+        self.patch_ln2 = nn.LayerNorm(mm_embed_dim)
+
+        # Factorized 2D positional embedding table: (mm_posemb_size, 2, mm_embed_dim)
+        self.pos_embedding = nn.Parameter(
+            torch.zeros(config.mm_posemb_size, 2, mm_embed_dim)
+        )
+        self.pos_norm = nn.LayerNorm(mm_embed_dim)
+
+    def forward(
+        self, pixel_values: torch.Tensor, image_position_ids: torch.Tensor
+    ) -> torch.Tensor:
+        # pixel_values: (B, num_patches, patch_dim); image_position_ids: (B, num_patches, 2)
+        hidden_states = self.patch_ln1(pixel_values.to(self.patch_dense.weight.dtype))
+        hidden_states = self.patch_dense(hidden_states)
+        hidden_states = self.patch_ln2(hidden_states)
+
+        clamped = image_position_ids.clamp(min=0).long()
+        valid = (image_position_ids != -1).to(self.pos_embedding.dtype).unsqueeze(-1)
+        axes = torch.arange(2, device=image_position_ids.device)
+        pos_embs = (self.pos_embedding[clamped, axes] * valid).sum(-2)
+        hidden_states = hidden_states + pos_embs
+        hidden_states = self.pos_norm(hidden_states)
+        return hidden_states
+
+
+class Gemma4UnifiedMultimodalEmbedder(nn.Module):
+    """Shared vision/audio projection: ``RMSNorm(no scale) -> Linear`` to LM space.
+
+    Both the vision and audio configs expose ``output_proj_dims`` (the projection
+    input dim) and ``rms_norm_eps``.  ``embedding_pre_projection_norm`` has no
+    learnable scale, so the only checkpoint tensor is ``embedding_projection.weight``.
+    """
+
+    def __init__(self, multimodal_config, text_config):
+        super().__init__()
+        self.multimodal_hidden_size = multimodal_config.output_proj_dims
+        self.text_hidden_size = text_config.hidden_size
+        self.embedding_pre_projection_norm = Gemma4RMSNorm(
+            self.multimodal_hidden_size,
+            eps=multimodal_config.rms_norm_eps,
+            with_scale=False,
+        )
+        self.embedding_projection = nn.Linear(
+            self.multimodal_hidden_size, self.text_hidden_size, bias=False
+        )
+
+    def forward(self, inputs_embeds: torch.Tensor) -> torch.Tensor:
+        inputs_embeds = inputs_embeds.to(self.embedding_projection.weight.dtype)
+        normed = self.embedding_pre_projection_norm(inputs_embeds)
+        return self.embedding_projection(normed)
+
+
+class Gemma4UnifiedForConditionalGeneration(Gemma4ForConditionalGeneration):
+    """Encoder-free unified Gemma4 (text + vision + audio).
+
+    Reuses the Gemma4 text decoder and the multimodal ``forward`` / attention
+    plumbing from :class:`Gemma4ForConditionalGeneration`, swapping the SigLIP
+    vision tower and conformer audio tower for the encoder-free embedders.
+    """
+
+    def __init__(
+        self,
+        config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        # Skip Gemma4ForConditionalGeneration.__init__ (it builds the SigLIP /
+        # conformer towers we do not have) and initialise the HF base directly.
+        PreTrainedModel.__init__(self, config=config)
+        self.pp_group = get_pp_group()
+        self.config = config
+        self.quant_config = quant_config
+
+        text_config = config.text_config
+
+        # Encoder-free embedders are consumed only at the input-embedding stage,
+        # so they live on the first PP rank only.
+        if self.pp_group.is_first_rank:
+            self.vision_embedder = (
+                Gemma4UnifiedVisionEmbedder(config.vision_config)
+                if getattr(config, "vision_config", None) is not None
+                else None
+            )
+            self.embed_vision = (
+                Gemma4UnifiedMultimodalEmbedder(config.vision_config, text_config)
+                if getattr(config, "vision_config", None) is not None
+                else None
+            )
+            self.embed_audio = (
+                Gemma4UnifiedMultimodalEmbedder(config.audio_config, text_config)
+                if getattr(config, "audio_config", None) is not None
+                else None
+            )
+        else:
+            self.vision_embedder = None
+            self.embed_vision = None
+            self.embed_audio = None
+
+        # Placeholders so methods inherited from the tower-based parent that
+        # reference these attributes never AttributeError.
+        self.vision_tower = None
+        self.audio_tower = None
+
+        self.vocab_size = text_config.vocab_size
+        self.vocab_size_per_layer_input = getattr(
+            text_config, "vocab_size_per_layer_input", text_config.vocab_size
+        )
+
+        self.language_model = Gemma4TextModel(
+            text_config,
+            quant_config,
+            prefix=add_prefix("language_model", add_prefix("model", prefix)),
+        )
+
+        text_tie = getattr(text_config, "tie_word_embeddings", True)
+        if self.pp_group.world_size == 1 and text_tie:
+            self.lm_head = self.language_model.embed_tokens
+        elif self.pp_group.is_last_rank:
+            self.lm_head = ParallelLMHead(
+                text_config.vocab_size,
+                text_config.hidden_size,
+                quant_config=quant_config,
+                prefix=add_prefix("lm_head", prefix),
+            )
+        else:
+            self.lm_head = PPMissingLayer()
+
+        self.logits_processor = LogitsProcessor(text_config)
+        self.capture_aux_hidden_states = False
+
+        # The unified checkpoint folds mm-projection vectors into the eoi/eoa
+        # rows of the (tied) embed_tokens, which inflates their lm-head logits.
+        # These are input-only markers that must never be sampled — HF applies a
+        # SuppressTokensLogitsProcessor for exactly these ids
+        # (generation_config.suppress_tokens). We reproduce that by masking only
+        # the next-token logits (input-logprob scoring of real eoi/eoa input
+        # tokens is left untouched).
+        suppress = []
+        for attr in ("eoi_token_id", "eoa_token_id", "eoa_token_index"):
+            tok_id = getattr(config, attr, None)
+            if isinstance(tok_id, int):
+                suppress.append(tok_id)
+        self.suppress_token_ids = sorted(set(suppress))
+        # Pre-materialize the index as a (non-persistent) buffer so it lives on
+        # the model's device — indexing with a Python list builds a CPU index
+        # tensor, which fails CUDA-graph capture ("cannot copy CPU<->CUDA").
+        self.register_buffer(
+            "_suppress_idx",
+            torch.tensor(self.suppress_token_ids, dtype=torch.long),
+            persistent=False,
+        )
+
+        self.post_init()
+
+    @torch.no_grad()
+    def forward(self, *args, **kwargs):
+        out = super().forward(*args, **kwargs)
+        if (
+            self._suppress_idx.numel() > 0
+            and isinstance(out, LogitsProcessorOutput)
+            and out.next_token_logits is not None
+        ):
+            out.next_token_logits.index_fill_(
+                1, self._suppress_idx, torch.finfo(out.next_token_logits.dtype).min
+            )
+        return out
+
+    # ------------------------------------------------------------------
+    # Per-modality feature extraction (encoder-free)
+    # ------------------------------------------------------------------
+    def _empty_embeds(self) -> torch.Tensor:
+        return torch.empty(
+            0,
+            self.language_model.config.hidden_size,
+            device=next(self.parameters()).device,
+            dtype=self.language_model.dtype(),
+        )
+
+    def _embed_patches(
+        self, items: List[MultimodalDataItem], position_attr: str
+    ) -> torch.Tensor:
+        all_embeds = []
+        for item in items:
+            all_pixel_values = flatten_nested_list([item.feature])
+            all_position_ids = flatten_nested_list([getattr(item, position_attr, None)])
+            for pv_idx, pv in enumerate(all_pixel_values):
+                # Pre-embedded passthrough (already at text hidden size).
+                if (
+                    pv.dim() in (2, 3)
+                    and pv.shape[-1] == self.config.text_config.hidden_size
+                ):
+                    all_embeds.append(pv.to(self.language_model.device))
+                    continue
+
+                if pv_idx >= len(all_position_ids) or all_position_ids[pv_idx] is None:
+                    raise ValueError(
+                        f"pixel_values[{pv_idx}] has no matching {position_attr}. "
+                        "The HF image/video processor likely renamed this output — "
+                        "update ATTR_NAME_TO_MODALITY in the Gemma4Unified processor."
+                    )
+                pp = all_position_ids[pv_idx]
+
+                # Collapse video (num_videos, num_frames, P, ...) -> (frames, P, ...)
+                if pv.dim() == 4:
+                    pv = pv.reshape(-1, pv.shape[-2], pv.shape[-1])
+                if pp.dim() == 4:
+                    pp = pp.reshape(-1, pp.shape[-2], pp.shape[-1])
+                if pv.dim() == 2:
+                    pv = pv.unsqueeze(0)
+                if pp.dim() == 2:
+                    pp = pp.unsqueeze(0)
+
+                pv = pv.to(
+                    device=self.language_model.device, dtype=self.language_model.dtype()
+                )
+                pp = pp.to(device=self.language_model.device)
+
+                embedded = self.vision_embedder(pv, pp)  # (B, P, mm_embed_dim)
+                projected = self.embed_vision(embedded)  # (B, P, hidden)
+
+                # Drop padding patches (position_ids == -1 on both axes).
+                padding_mask = (pp == -1).all(dim=-1)  # (B, P)
+                all_embeds.append(projected[~padding_mask])
+
+        return torch.cat(all_embeds, dim=0) if all_embeds else self._empty_embeds()
+
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        return self._embed_patches(items, "image_position_ids")
+
+    def get_video_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        return self._embed_patches(items, "video_position_ids")
+
+    def get_audio_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        if self.embed_audio is None:
+            raise ValueError(
+                "Audio inputs provided but the model was built without an audio_config."
+            )
+        all_input_features = flatten_nested_list([item.feature for item in items])
+        # input_features_mask convention: True = valid token.
+        all_masks = flatten_nested_list([item.input_features_mask for item in items])
+
+        all_embeds = []
+        for input_features, mask in zip(all_input_features, all_masks):
+            if input_features.dim() == 2:
+                input_features = input_features.unsqueeze(0)
+            if mask.dim() == 1:
+                mask = mask.unsqueeze(0)
+            input_features = input_features.to(
+                device=self.language_model.device, dtype=self.language_model.dtype()
+            )
+            mask = mask.to(device=input_features.device)
+
+            # Raw waveform frames -> RMSNorm -> Linear (no conformer/mel).
+            projected = self.embed_audio(inputs_embeds=input_features)  # (B, T, hidden)
+            for enc, m in zip(projected, mask):
+                all_embeds.append(enc[m])  # keep valid frames only
+
+        return torch.cat(all_embeds, dim=0) if all_embeds else self._empty_embeds()
+
+    # ------------------------------------------------------------------
+    # Weight loading
+    # ------------------------------------------------------------------
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]) -> Set[str]:
+        k_eq_v_layers = self._get_k_eq_v_layers()
+
+        params_dict = dict(self.named_parameters())
+        params_dict.update(dict(self.named_buffers()))
+        non_persistent_buffers: Set[str] = set()
+        for mod_name, mod in self.named_modules():
+            for buf_name in getattr(mod, "_non_persistent_buffers_set", set()):
+                full = f"{mod_name}.{buf_name}" if mod_name else buf_name
+                non_persistent_buffers.add(full)
+
+        text_tie = getattr(self.config.text_config, "tie_word_embeddings", True)
+        start_layer = self.language_model.start_layer
+        end_layer = self.language_model.end_layer
+
+        loaded_params: Set[str] = set()
+
+        for name, loaded_weight in weights:
+            name = re.sub(r"^model\.", "", name)
+
+            if pp_filter_load_weight(
+                name,
+                loaded_weight,
+                pp_group=self.pp_group,
+                start_layer=start_layer,
+                end_layer=end_layer,
+                params_dict=params_dict,
+                loaded_params=loaded_params,
+                tie_word_embeddings=text_tie,
+                embed_weight_name="language_model.embed_tokens.weight",
+                first_rank_only_patterns=(
+                    "language_model.embed_tokens",
+                    "language_model.per_layer_model_projection",
+                    "language_model.per_layer_projection_norm",
+                    "vision_embedder.",
+                    "embed_vision.",
+                    "embed_audio.",
+                ),
+                last_rank_only_prefixes=("language_model.norm.", "lm_head."),
+            ):
+                continue
+
+            # attention_k_eq_v: full-attention layers ship only k_proj (V == K).
+            # Load k_proj into both the "k" and "v" shards of the fused QKV.
+            should_dup_k_to_v = (
+                ".k_proj." in name
+                and k_eq_v_layers
+                and "language_model." in name
+                and (m := re.search(r"layers\.(\d+)\.", name)) is not None
+                and int(m.group(1)) in k_eq_v_layers
+            )
+
+            for param_name, weight_name, shard_id in self.stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                mapped = name.replace(weight_name, param_name)
+                if mapped not in params_dict:
+                    continue
+                param = params_dict[mapped]
+                param.weight_loader(param, loaded_weight, shard_id)
+                if should_dup_k_to_v:
+                    param.weight_loader(param, loaded_weight, "v")
+                loaded_params.add(mapped)
+                break
+            else:
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+                loaded_params.add(name)
+
+        unloaded_params = params_dict.keys() - loaded_params
+        if unloaded_params:
+            param_names = set(dict(self.named_parameters()).keys())
+            buckets = {
+                logging.WARNING: (
+                    "Some weights are not initialized from checkpoints",
+                    lambda p: p in param_names,
+                ),
+                logging.INFO: (
+                    "Persistent buffers not in checkpoint (using default init)",
+                    lambda p: p not in param_names and p not in non_persistent_buffers,
+                ),
+                logging.DEBUG: (
+                    "Non-persistent buffers not in checkpoint (expected)",
+                    lambda p: p in non_persistent_buffers,
+                ),
+            }
+            for level, (msg, pred) in buckets.items():
+                names = sorted(p for p in unloaded_params if pred(p))
+                if names:
+                    logger.log(level, "%s: %s", msg, names)
+        return loaded_params
+
+
+EntryClass = Gemma4UnifiedForConditionalGeneration
@@ -416,6 +416,7 @@ class BaseMultimodalProcessor(ABC):
            if self._processor.__class__.__name__ in {
                "Gemma3nProcessor",
                "Gemma4Processor",
+                "Gemma4UnifiedProcessor",
                "GlmAsrProcessor",
                "Qwen2AudioProcessor",
                "Qwen3ASRProcessor",
@@ -0,0 +1,33 @@
+# Copyright 2026 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+from sglang.srt.models.gemma4_unified import Gemma4UnifiedForConditionalGeneration
+from sglang.srt.multimodal.processors.gemma4 import Gemma4SGLangProcessor
+
+
+class Gemma4UnifiedSGLangProcessor(Gemma4SGLangProcessor):
+    """Multimodal processor for the encoder-free unified Gemma4 (image/video/audio).
+
+    Identical to :class:`Gemma4SGLangProcessor` except for audio padding: the
+    unified model has no SSCP conformer, so the waveform is simply chunked into
+    fixed ``audio_samples_per_token`` (640) frames.  Padding the waveform up to a
+    multiple of that frame size keeps ``ceil(num_samples / spt)`` consistent with
+    the number of valid frames the feature extractor emits.
+    """
+
+    models = [Gemma4UnifiedForConditionalGeneration]
+
+    def _get_audio_pad_multiple(self) -> int:
+        fe = getattr(self._processor, "feature_extractor", None)
+        return getattr(fe, "audio_samples_per_token", 640)
@@ -2332,6 +2332,7 @@ class ServerArgs:
        elif model_arch in (
            "Gemma4ForConditionalGeneration",
            "Gemma4ForCausalLM",
+            "Gemma4UnifiedForConditionalGeneration",
        ):
            default_attention_backend = (
                "trtllm_mha" if is_sm100_supported() else "triton"
@@ -133,7 +133,12 @@ class HfModelConfigParser(ModelConfigParserBase):
        if config.model_type == "multi_modality":
            _set_architectures(config, "MultiModalityCausalLM")

-        if config.model_type in ("gemma4", "gemma4_assistant"):
+        if config.model_type in (
+            "gemma4",
+            "gemma4_assistant",
+            "gemma4_unified",
+            "gemma4_unified_assistant",
+        ):
            # Gemma4 configs use base attributes for SWA layers and `global_*`
            # variants for full-attention layers.  SGLang expects the opposite:
            # base = full-attention, `swa_*` = sliding-window overrides.
@@ -158,6 +163,13 @@ class HfModelConfigParser(ModelConfigParserBase):
            if not hasattr(text_config, "swa_v_head_dim"):
                text_config.swa_v_head_dim = text_config.swa_head_dim

+            # Unified Gemma4 names the end-of-audio token `eoa_token_index`,
+            # but the multimodal processor expects `eoa_token_id`.
+            if not hasattr(config, "eoa_token_id") and hasattr(
+                config, "eoa_token_index"
+            ):
+                config.eoa_token_id = config.eoa_token_index
+
        if config.model_type == "longcat_flash":
            _set_architectures(config, "LongcatFlashForCausalLM")