add: update scripts for gemma4 26B-A4B testing

[jungledesh]

Greptile Summary

The PR aligns bundled launch and load scripts around explicit model-family and served-name settings, and adds NVIDIA/AMD Gemma 4 launch support. The AMD image-level Llama name currently leaks into the Gemma launcher and prevents correct model identification.

• Makes Gemma 4 26B-A4B the NVIDIA default and keeps Llama 3 as the AMD default.
• Adds served-name-aware load and swarm configuration.
• Adds an AMD Gemma launcher and a vendored Gemma tool-call template.

Confidence Score: 4/5

The PR should not merge until the AMD Gemma switch stops inheriting the Llama served name and preserves correct Gemma model identification.

The AMD image exports SERVED_NAME=llama3, and the new Gemma launcher only supplies its Gemma name when that variable is unset, so the documented switch serves Gemma weights under a Llama identifier and deprives Profile of the correct model metadata.

Files Needing Attention: Dockerfile, scripts/start-gemma-amd.sh

Important Files Changed

Filename	Overview
Dockerfile	Adds model-specific defaults and launchers, but the AMD-wide Llama served name overrides the documented Gemma launcher default.
scripts/start-gemma-amd.sh	Adds AMD Gemma startup and tool support, while preserving an incompatible image-level served name when invoked as documented.
scripts/load.sh	Correctly resolves request names in MODEL, SERVED_NAME, then model-family order and fixes the previously reported default mismatches.
scripts/agent-swarm.sh	Adds Gemma/Qwen-specific aliases and consistently propagates the resolved served model name.
scripts/start-gemma.sh	Updates NVIDIA startup for Gemma 4 26B-A4B and adds the required tool-call parser and template configuration.
scripts/start-amd.sh	Makes the AMD Llama served name overridable and consistent with load generation.
scripts/tool_chat_template_gemma4.jinja	Adds the Gemma tool-call chat template used by the new launchers.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[AMD image ENV<br/>SERVED_NAME=llama3] --> B[start-gemma.sh]
  B --> C[start-gemma-amd.sh preserves inherited name]
  C --> D[vLLM serves Gemma weights as llama3]
  D --> E[Profile cannot select Gemma metadata]

Loading

Prompt To Fix All With AI

### Issue 1
Dockerfile:198
**AMD Gemma inherits Llama identity**

When the AMD container starts `./start-gemma.sh` as documented without overriding `SERVED_NAME`, the image-level `llama3` value prevents the launcher from applying its Gemma default, causing Gemma weights to be advertised under a Llama identifier and Profile to omit the correct Gemma-specific diagnostics.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (5): Last reviewed commit: "script in dockerfile" | Re-trigger Greptile}

Greptile also left 1 inline comment on this PR.

[greptile-apps]

AMD Gemma inherits Llama identity

When the AMD container starts ./start-gemma.sh as documented without overriding SERVED_NAME, the image-level llama3 value prevents the launcher from applying its Gemma default, causing Gemma weights to be advertised under a Llama identifier and Profile to omit the correct Gemma-specific diagnostics.

Prompt To Fix With AI

This is a comment left during a code review.
Path: Dockerfile
Line: 198

Comment:
**AMD Gemma inherits Llama identity**

When the AMD container starts `./start-gemma.sh` as documented without overriding `SERVED_NAME`, the image-level `llama3` value prevents the launcher from applying its Gemma default, causing Gemma weights to be advertised under a Llama identifier and Profile to omit the correct Gemma-specific diagnostics.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.