DSpark support model does not match the pinned DeepSeek V4 Flash Vision-Exp checkpoint

[Charlie-Louis]

Tittle says it all.

The drafter cannot be used with Vision-Exp

./ds4-server --model "$MODEL" --vision "$VISION" --metal --ctx 250000 --host 0.0.0.0 --mtp-model ../gguf/DeepSeek-V4-Flash-DSpark-support-0731.gguf --dspark
ds4: DSpark support model does not match the pinned DeepSeek V4 Flash Vision-Exp checkpoint

The default downloader script takes the none vision drafter, should check or ask for option

[gilbert-barajas]

Sharing a header-level read that may save others the same detour (M5 Max, Metal, Vision-Exp IQ2XXS build, engine 110afdd):

Answer at the header/loader level: (a), with a metadata pin — not (c)

The pin is metadata, not tensor shapes. When the target is Vision-Exp
(g_ds4_flash_vision_exp, set at ds4.c:5919 from the target's own
deepseek4.checkpoint_variant / general.source.revision), the loader requires
the support GGUF to carry BOTH strings (ds4.c:2922
support_model_checkpoint_compatible):

That is the exact refusal #949 reports ("does not match the pinned … Vision-Exp
checkpoint"). The text-Flash DSpark-support-0731 support model does not carry
those strings, so it is refused — correctly. Nothing about the refusal
implies shape incompatibility; it is a provenance pin, and the loader's shape
checks (dspark_weights_bind_optional → dspark_weights_validate_layout) come
later with different messages.

The matching artifact already exists upstream and passes the pin. Shipped
in the same repo the vision model comes from:

./download_model.sh ds4f-vision-dspark     # antirez/deepseek-v4-gguf
# → DeepSeek-V4-Flash-Vision-Exp-DSpark-support.gguf, 5,989,114,528 bytes
#   sha256 0807a67fd9ce5874bfc60d8d2461f50e11657e3dd94913d3473f85aa679bc877

Header fields (read directly from the GGUF, GGUF v3, 81 tensors, 12 KVs):

field	value
general.architecture	deepseek4-dspark
general.name	DeepSeek V4 Flash Vision Experimental DSpark support
deepseek4.checkpoint_variant	vision-exp ← passes the pin
general.source.revision	e46e16bf6035c6f317eb2ac7458eb0362926d402 ← passes the pin
dspark.stage_count	3
dspark.block_size	5
dspark.markov_rank	256
dspark.noise_token_id	128799 (< target vocab 129280 ✓)
dspark.target_layer_ids	[40, 41, 42]

Target header, for the shape comparison (#949 asked "does the Flash DSpark head
fit the Vision-Exp checkpoint?"):

field	Vision-Exp target
general.architecture	deepseek4
deepseek4.checkpoint_variant	vision-exp
general.source.revision	e46e16bf6035c6f317eb2ac7458eb0362926d402
deepseek4.embedding_length	4096
deepseek4.vocab_size	129280
deepseek4.attention.head_count / key_length	64 / 512
deepseek4.block_count	43
deepseek4.expert_count / expert_used_count	256 / 6

So for #948: at the header level the artifact asked for already exists in antirez/deepseek-v4-gguf (./download_model.sh ds4f-vision-dspark). Acceptance/speed numbers on this box to follow once we have a clean measurement; happy to run any specific config the maintainers want to see.