deepseek4: NextN/MTP speculative decoding (converter + graph)#25642
Closed
bownux wants to merge 3 commits into
Closed
deepseek4: NextN/MTP speculative decoding (converter + graph)#25642bownux wants to merge 3 commits into
bownux wants to merge 3 commits into
Conversation
- gguf-py: add NEXTN_HC_HEAD_{FN,BASE,SCALE} tensor types; register the 7
nextn tensors in DEEPSEEK4's MODEL_TENSORS
- conversion/deepseek.py: rekey mtp.{i}.* -> layers.{n_layer+i}.* at init so
FP8 pairing / expert merge / MXFP4 repack apply unchanged; map the glue
tensors; concat e_proj+h_proj -> nextn.eh_proj {2*n_embd, n_embd}
(eh_proj(concat[e;x]) == e_proj(e) + h_proj(x)); include NextN layers in
the MXFP4 expert loop
Verified: 1,359-tensor GGUF with all 7 glue tensors + full blk.43 layer.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01RNCV8Tr1GAhQMomrZdS676
First working DeepSeek-V4 MTP on llama.cpp. The NextN block is a full V4
layer (pure-SWA MLA, MoE, hyper-connections) per the reference MTPBlock:
logits = shared_head(shared_head_norm(hc_head(V4Block(
eh_proj(concat[enorm(embed(tok)); hnorm(h)])))))
- deepseek4: read nextn_predict_layers FIRST (n_layer() sizes hparam arrays);
mark nextn layers SWA + compress_ratio 0; n_embd_nextn = hc_mult*n_embd
(the MTP block exchanges the full hyper-connection state)
- loader: extend the per-layer loop to n_layer_all; load the 7 nextn glue
tensors (eh_proj packs e_proj|h_proj via the concat identity)
- graph_mtp: qwen35-pattern draft graph reusing the trunk's hc/attention/moe
builders at il = n_layer; t_h_nextn = raw hc state for chaining
- trunk graph: expose t_h_nextn (pre-hc_head, all tokens)
- hparams/context: n_embd_nextn() accessor; nextn buffer sizing/indexing and
MTP batch-embd width honor the wider rows
- speculative: draft-mtp width from llama_model_n_embd_nextn (new ext API);
verify_h row stats at trace verbosity
- kv-cache: deepseek4 MTP-context layer filter (STEP35 pattern); fix SIGFPE
in build_input_k_rot when a layer filter selects no layers
- arch: NEXTN_HC_HEAD_{FN,BASE,SCALE} tensor types
- DSV4_MTP_DEBUG=1 env for load/graph debug logging
Validated: 44-layer MTP GGUF loads; trunk unchanged; draft-mtp generates
coherent output with 82% draft acceptance (172 drafted / 141 accepted),
deterministic across runs.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01RNCV8Tr1GAhQMomrZdS676
…les)
llama-quantize aborted on GGML_TYPE_I32 tensors ('cannot dequantize/convert
tensor type i32') because the Q-mix tried to requantize them. Integer tensors
are lookup/routing tables, never weights - skip them.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01RNCV8Tr1GAhQMomrZdS676
github-actions
Bot
added
documentation
|
Hi @bownux, thanks for your contribution! Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:
Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below. |
Contributor
|
I will be adding this in the future. It requires some more changes |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Implements DeepSeek-V4's multi-token-prediction (NextN/MTP) head end to end for the
deepseek4arch, so V4 GGUFs can use--spec-type draft-mtplike qwen35/step35/hy_v3.What
conversion/deepseek.py): keep themtp.*tensors (currently dropped) by rekeying them as trailing layern_layerat ingestion, so the existing FP8/MXFP4/expert-merge machinery applies unchanged. DeepSeek'se_proj/h_projglue folds into the existingnextn.eh_projconvention viae_proj(e) + h_proj(x) = eh_proj(concat[e; x]). Writesblock_count = n_layer + nextnandnextn_predict_layers, matching the step35 convention.src/models/deepseek4.cpp):graph_mtpfor the NextN block - enorm/hnorm -> eh_proj -> a full V4 layer (MLA attention in pure-SWA mode, 256-expert MoE, hyper-connection mixers) -> shared head. V4's hyper-connections make the drafter's hidden-state inputhc_mult * n_embd(4x) wide; plumbed asn_embd_nextn()through llama-context and the speculative driver.Included fixes (hit while testing; can split into separate PRs on request)
llama-quant.cpp: pass integer tensors (I8/I16/I32/I64) through unquantized - deepseek4'stid2eidrouting tables crashedllama-quantize.llama-kv-cache.cpp: guardbuild_input_k_rotagainstn_embd_head_k_all == 0(SIGFPE / infinite loop when a KV-cache filter leaves a cache empty).Testing
On 4x RTX PRO 4500 + Threadripper 9970X, DeepSeek-V4-Flash converted with this branch (Q8/native-MXFP4 and Q3_K_M requant, both public at https://huggingface.co/rogerai-fyi/DeepSeek-V4-Flash-MTP-GGUF):
MTPBlock(inference/model.py)--spec-draft-n-max 1: 100% draft acceptance, 23.5 -> 30.5 tok/s (+30%) on the Q3 quant with 10 expert layers on CPU; acceptance 93%/78% at n_max 2/3 as expectedDraft-quality note: models converted with earlier mainline (which dropped
mtp.*) simply have no NextN tensors and are unaffected; this only adds capability for newly-converted GGUFs.🤖 Generated with Claude Code
https://claude.ai/code/session_01RNCV8Tr1GAhQMomrZdS676