feat(laguna): add Domino GRU draft head

[davide221]

Summary

• embed optional Domino GRU/head aux tensors into DFlash draft GGUFs via convert_dflash_to_gguf.py
• load and validate dflash.domino.* tensors in the draft GGUF loader
• add a Laguna target full-logits projection hook and greedy-chain Domino correction path
• keep DDTree and sampled verify on the existing base projection path for now

Bench notes

Ran the 30-case he,gsm,math harness on lucebox-ts RTX 3090 with Laguna-XS.2 target, max_ctx 4096, default cache, no DDTree.

v13-final+Domino:

• HumanEval: 10/10, 172.94 tok/s
• GSM8K: 7/10, 155.27 tok/s
• MATH: 8/10, 149.53 tok/s
• server aggregate: 151.56 weighted tok/s, 60.3% accept, 3.03 avg commit

Compared with v12, v13-final+Domino is slower but improves GSM by one case. This PR is the runtime/model-loading support, not a claim that v13 is the best drafter.

Checks

• python3 -m py_compile server/scripts/convert_dflash_to_gguf.py
• git diff --check
• cmake --build server/build -j -t dflash_server on lucebox-ts existing CUDA build

Note: a fresh clean-worktree CUDA rebuild on the 3090 was attempted, but the full llama.cpp CUDA dependency compile was killed with Error 137 under memory pressure before reaching the project-only compile.

[cubic-dev-ai]

1 issue found across 11 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="server/src/draft/draft_gguf_loader.cpp">

<violation number="1" location="server/src/draft/draft_gguf_loader.cpp:119">
P2: Domino 1D shape validation is incomplete: it only checks `ne[0]` and does not enforce that higher dimensions are 1, so malformed non-vector tensors can pass loader validation.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: Domino 1D shape validation is incomplete: it only checks ne[0] and does not enforce that higher dimensions are 1, so malformed non-vector tensors can pass loader validation.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At server/src/draft/draft_gguf_loader.cpp, line 119:

<comment>Domino 1D shape validation is incomplete: it only checks `ne[0]` and does not enforce that higher dimensions are 1, so malformed non-vector tensors can pass loader validation.</comment>

<file context>
@@ -115,6 +115,29 @@ int count_swa_layers(const DraftWeights & w) {
 }
 
+bool check_shape_1d(const ggml_tensor * t, int64_t ne0, const char * name, char * buf, size_t buf_sz) {
+    if (!t || t->ne[0] != ne0) {
+        std::snprintf(buf, buf_sz, "draft GGUF: Domino tensor %s shape mismatch: got [%lld], expected [%lld]",
+                      name, t ? (long long)t->ne[0] : -1LL, (long long)ne0);
</file context>

Suggested change

	if (!t || t->ne[0] != ne0) {
	if (!t || t->ne[0] != ne0 || t->ne[1] != 1 || t->ne[2] != 1 || t->ne[3] != 1) {