dflash: load and apply DSpark log-SNR conditioning

[bri-prism]

Re-lands the DSpark log-SNR conditioning from the closed prism-v6 work, which is what currently makes GIDD-bundle drafters unloadable on this branch.

The symptom

llama_model_load: error loading model: done_getting_tensors: wrong number of tensors; expected 79, got 75
common_speculative_init_result: failed to load draft model

The message names no tensor, which is misleading. The file has 79 tensors and the loader creates 75, and the natural first guess is the DSpark Markov head, since that is the one optional block in load_arch_tensors. It is not: instrumenting the loader shows markov_w1.weight is found and the head's four tensors are created normally.

The four that are never created are log_snr_fc1.weight, log_snr_fc1.bias, log_snr_fc2.weight and log_snr_fc2.bias. This branch has no log-SNR support at all, so nothing ever claims them, and done_getting_tensors refuses the load.

Why not just skip them

The drafter's metadata sets log_snr_conditioning = true along with min_log_snr and max_log_snr. That means it was trained with the conditioning, so the weights are part of the function. Marking the tensors optional would let the model load and quietly compute a different draft embedding on every forward pass. A load failure is the better of the two, which is why this ports the feature rather than relaxing the check.

What this adds

• three metadata keys, two tensor kinds, the hparams, and the four model tensors
• llm_graph_input_dspark_logsnr, which stages the sinusoidal feature matrix. The per-position log-SNR pattern (each block's anchor at max_log_snr, mask positions at min_log_snr) and its featurization are a pure function of values known at graph-build time, so the host precomputes the whole [128, n_tokens] matrix once. That keeps it auditable line-for-line against the training reference instead of chaining arange/sin/cos in-graph.
• the two-layer SiLU MLP, added to the draft noise embedding before the layer loop

The weights are required when the flag is set, and min/max must be finite with max > min, because a zero span divides through the featurizer and fills the conditioning input with NaNs rather than failing.

The DSV4 hyper-connection backbone replicates the embedding across hc lanes, so the term would need adding per lane. No checkpoint exercises that combination yet, so that path asserts rather than silently dropping a trained input.

Validation

Measured on Metal with a block-4 drafter that sets the flag, against a low-bit target:

	result
before	drafter cannot be loaded
after	65.891% acceptance, mean accepted length 3.58

The same numbers come out with this branch alone, and with this branch plus the layout auto-detect in the sibling PR, on both the correct and the deliberately wrong --spec-type. That last case matters because the auto-detect could not be exercised on this branch at all before this change, since no DSpark drafter would load.

Not covered: CUDA. This is graph and loader code with no backend-specific paths, and it builds clean, but every number above is Metal.

[khosravipasha]

This looks clean merging first.

[Copilot]

Pull request overview

Adds DSpark log-SNR conditioning so GIDD-bundle drafters load and apply their trained conditioning.

Changes:

• Adds GGUF metadata and tensor definitions.
• Validates and loads conditioning parameters and weights.
• Applies the sinusoidal two-layer conditioning MLP in DFlash graphs.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/models/dflash.cpp	Loads, validates, and applies conditioning.
src/llama-model.h	Stores conditioning tensors.
src/llama-hparams.h	Stores conditioning metadata.
src/llama-graph.h	Defines the graph input.
src/llama-graph.cpp	Uploads precomputed features.
src/llama-arch.h	Adds metadata and tensor enums.
src/llama-arch.cpp	Registers metadata and tensor names.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[bri-prism]

Good catch, and this one was a real regression rather than a style point. Fixed in 9ed9b9b.

You are right that allow_reuse already compares n_tokens and n_seqs_unq, which together with min/max_log_snr (fixed per model) are everything v_feat depends on, so the input can opt in. can_reuse now returns true when the staged tensor still matches n_tokens, following the same shape as llm_graph_input_embd::can_reuse.

Re-ran the drafter afterwards to confirm nothing moved: acceptance 65.891%, mean length 3.58, identical to before.