DeepSeek-V4-Pro is the same architecture as V4-Flash: record the variant and prove the config descent is shape-generic

[localai-bot]

deepseek-ai/DeepSeek-V4-Pro is DeepseekV4ForCausalLM / model_type: deepseek_v4 — the same architecture as DeepSeek-V4-Flash, which we already support (MODEL-TEXT-deepseek-v4-deepseek-v4-for-causal-lm, ✅). This issue records the assessment and adds the evidence that our config descent is genuinely shape-generic rather than Flash-tuned.

Grounding

Config diff (key-by-key, both config.json from HF): zero new keys, zero removed keys. Every difference is a scaled value.

key	Flash	Pro
hidden_size	4096	7168
num_hidden_layers	43	61
num_attention_heads	64	128
n_routed_experts	256	384
moe_intermediate_size	2048	3072
q_lora_rank	1024	1536
o_groups	8	16
index_topk	512	1024
routed_scaling_factor	1.5	2.5
compress_ratios[0:2]	[0, 0, …]	[128, 128, …]

Identical on both: head_dim 512, qk_rope_head_dim 64, hc_mult 4, hc_eps, hc_sinkhorn_iters 20, num_hash_layers 3, n_shared_experts 1, sliding_window 128, scoring_func: sqrtsoftplus, topk_method: noaux_tc, swiglu_limit 10.0, expert_dtype: fp4, index_head_dim 128, index_n_heads 64, compress_rope_theta 160000, num_experts_per_tok 6, vocab_size 129280. Heads-per-output-LoRA-group is 8 on both (64/8, 128/16).

Upstream (pinned oracle 5559679229bc961848b121ccdeaa8fa5d79bec98, vLLM 0.26.0.dev0): the same vllm/models/deepseek_v4/ package serves both. Every dimension is read off config (nvidia/model.py:535-679, attention.py:193-211, attention.py:708). Compressor presence is compress_ratio > 1 (attention.py:334) over max(1, config.compress_ratios[layer_id]) (attention.py:209), and the compressor itself asserts compress_ratio in [4, 128] (compressor.py:171). Pro is inside that asserted set. There is no Pro-specific code path upstream.

Our side: the only shape assertion across the 6930 lines of V4 sources is head_dim == 512 (deepseek_v4_weights.cpp:173), which Pro satisfies. All dims come off the config (deepseek_v4.cpp:1725-1731), all CUDA shared memory is dynamic (extern __shared__), and compress_ratios is consumed by value — has_compressor(l) = ratio != 0, has_indexer(l) = ratio == 4 (deepseek_v4.h:127-128), which matches upstream's > 1 / == 4 polarity exactly (ratio 0 → upstream max(1,0) = 1 → no compressor).

Checkpoint-level grounding. Normalizing layer and expert indices out of both real model.safetensors.index.json files gives 98 distinct tensor-name patterns on each side, with zero patterns unique to either. Predicting compressor/indexer layers from compress_ratios alone reproduces both real checkpoints exactly:

	predicted compressor / indexer	actual in checkpoint
Flash (43L)	41 / 21	41 / 21
Pro (61L)	61 / 30	61 / 30

Pro's layers 0-1 carry the 4-tensor ratio-128 compressor group where Flash has none, exactly as the value-driven rule predicts.

Verdict

No architecture work is owed. Pro is a config variant of the existing row, not a new model row.

What blocks actually running it

1.599T params. deepseek-ai/DeepSeek-V4-Pro is 805 GiB, nvidia/DeepSeek-V4-Pro-NVFP4 864 GiB, and the smallest real GGUF (teamblobfish/DeepSeek-V4-Pro-GGUF) is Q2_K-XL at 534.7 GiB. One GB10 is 119 GiB unified — Q2 is ~4.5x over, and 1.6T params at 2 bits has a ~400 GB floor. The pinned oracle needs the same memory, so there is no gateable denominator on this hardware either (Flash already needed 2 Sparks, and this needs roughly 5-8 at Q2).

Second, independent of memory: the real-geometry DSA sparse path is an already-named residual on our side. dsa_dense = (be.gguf != nullptr) (deepseek_v4.cpp:668) forces the compressor and indexer off on the real keep-quant path, so Flash runs dense MLA, which is exact only while seq_len <= index_topk. For Flash that window is 512 tokens; for Pro it is 1024. Long-context V4 on either model needs that residual built out.

Scope of this issue

Records-only plus one test. No download, no GPU, no forward:

• Add a CPU test that drives ParseDeepseekV4Params with the real Pro config values and asserts the derived geometry (compressor/indexer layer counts 61/30, nope 448, heads-per-group 8, zdim, expert count), with Flash's 43L/41/21 as the control arm.
• RED evidence by mutation: hardcoding a Flash dimension must turn the new test red.
• Record the Pro variant + the HW verdict in the row spec.

Out of scope: any attempt to load or run Pro, and the DSA real-geometry residual (tracked separately).