DiffusionGemma Ecosystem Status
Tracking PR merge status across llama.cpp, vLLM, and related projects. Updated when key events happen — last checked July 4, 2026. Both open PRs now mergeable — 1 of 3 PRs merged, 2 ready to land.
llama.cpp #24423 still in draft but now mergeable (conflict resolved, activity July 4). vLLM #46177 merged June 26 — the only PR that has landed. vLLM #46454 is ready and mergeable (rebase fixed the conflict; author confirmed ready June 27). Both open PRs are technically clear to merge. llama.cpp ecosystem (Ollama, LM Studio) still waiting on #24423 to exit draft.
llama.cpp #24423 — DiffusionGemma ⚠️ Draft
The main PR bringing DiffusionGemma support to llama.cpp — still in draft, but now mergeable (conflict resolved). Adds the diffusion architecture, GGUF conversion, and CUDA kernel support. Initially submitted as a proof-of-concept; maintainer feedback requested a more general diffusion server design. Must build from the PR branch — not available in mainline llama.cpp yet. Ollama and LM Studio cannot bundle until this lands.
Not merged. Maintainer (pwilkin) wants a general diffusion-server instead of model-specific. Conflict resolved — technically mergeable, pending design direction from @ggerganov. Last activity July 4, 2026.
vLLM #46177 — Tensor Parallelism Fix ✅ Merged
Fixes DiffusionGemma crashes on multi-GPU setups by adding tensor parallelism support. Merged June 26, 2026. Multi-GPU DiffusionGemma is now available in vLLM main.
First of the three to land. Independent validation from @shubhamprshr27 confirmed the TP=8 fix. Build from vLLM main to get multi-GPU support now.
vLLM #46454 — GGUF Tied Embedding Fix
Fixes GGUF quantized model loading for DiffusionGemma — tied embedding weights were breaking startup. Affects anyone running quantized models through vLLM.
Ready for review. #46177 merged — this PR is standalone. Mergify flagged a conflict June 26; author rebased and confirmed "ready" June 27. Now mergeable — no blockers.
What Each PR Unlocks
| PR | Unlocks | Current Workaround |
|---|---|---|
| llama.cpp #24423 ⚠️ | llama.cpp native, GGUF, Ollama, LM Studio | Build from PR branch — not in mainline |
| vLLM #46177 ✅ | Multi-GPU inference (dual 3090/4090) | Merged — build from vLLM main |
| vLLM #46454 | GGUF models in vLLM | Use BF16 weights or wait for merge |
Ecosystem Maturity
| Tool | Status | Notes |
|---|---|---|
| vLLM | Day-zero support | Single + multi-GPU. #46177 merged June 26 — tensor parallelism now in main. |
| llama.cpp | PR in draft (not merged) | #24423 is a proof-of-concept, still in draft with conflicts. Build from PR branch for now. |
| GGUF | PR branch only | Q4_K_M available on HuggingFace. Convert requires PR branch — not yet in mainline llama.cpp. |
| Unsloth | Day-zero support | 4-bit loading supported. VRAM table at unsloth.ai/docs. |
| HuggingFace | Day-zero support | trust_remote_code=True required for custom diffusion modules. |
| Ollama | Waiting for llama.cpp PR | #24423 not yet merged. Ollama cannot bundle until it lands in llama.cpp mainline. Bridge workaround still works. |
| LM Studio | Waiting for llama.cpp PR | #24423 not yet merged. LM Studio cannot bundle until it lands in llama.cpp mainline. Alternatives available. |
Save 40+ Hours of Debugging →
The free tutorials get you running. The Production Manual keeps you running — Docker Compose, GPU tuning, 20+ error fixes, real benchmarks. $39 $29, lifetime updates, 30-day refund.