model: add MTP support for Nemotron model by ruixiang63 · Pull Request #26725 · ggml-org/llama.cpp

ruixiang63 · 2026-08-07T13:46:31Z

Overview

This PR adds MTP support for the Nemotron Nano model. Performance is not yet optimal and depends on PR #26623 being merged first.
I tested this PR on the top of PR #26623, and the performance looks good.

Additional information

Requirements

I have read and agree with the contributing guidelines
AI usage disclosure: yes, paired with cursor

am17an

cpp side changes look ok.

gaugarg-nv · 2026-08-08T08:04:36Z

+                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        cb(cur, "mtp_attn_out", il);
+    }
+


The inp_out_ids gather can happen at this line, before the MoE FFN.
We need only attention to run over all n_tokens to fill KV.

This should help improve prefill perf a bit.

Makes sense. I moved the inp_out_ids gather before the MoE FFN so attention still runs over all tokens while the MoE only processes the kept rows.
Guarded it the same way mimo2 does (crop_before_ffn = inp_out_ids && (!embeddings_nextn || embeddings_nextn_masked)).

…, multi-output backend sampling, pocket-tts Merges 45 upstream commits (0865990..ebb546b, b10362-16-gebb546b7e) into synori/llama-update-mtp-fit. Zero conflicts; all vendored patches carried over untouched. Primary motivation — new Meta architecture: * 62bf73d model: Muse Glimmer Support (ggml-org#26841) LLM_ARCH_MUSE_GLIMMER + src/models/muse-glimmer.cpp + the mtmd vision tower in tools/mtmd/models/muse-glimmer.cpp and conversion/muse_glimmer.py. Other notable changes that touch our public API surface: * dd1ea52 llama : support multi-output backend sampling (ggml-org#25532) llama_context_params gains n_outputs_max_per_seq; llama_sampler_i.backend_init takes it as a third argument; new backend_reset / copy_state vtable slots and llama_sampler_copy(). * 153d324 llama : default load-mode auto, avoids mmap on iGPUs (ggml-org#26081) llama_load_mode gains LLAMA_LOAD_MODE_AUTO = -1 (enum is now signed). * 6e62ba5 mtmd: support pocket-tts (ggml-org#26871) mtmd_gen_inp/mtmd_gen_out gain seed/temp/feats/is_eos; new mtmd_gen_inp_default(); mtmd_helper_gen_audio_step_gen() takes out_stop. * 157b81f model : Granite-Switch Architecture (ggml-org#25107) * 7a20b41 model: MTP support for Nemotron (ggml-org#26725) and cc078b4 Dflash support for nemotron-3.5 (ggml-org#26905) * e23e944 vendor : cpp-httplib 0.53.0, 4c6766f vendor : subprocess.h sync Vendored patches preserved: * 919fde3 feat(rpc): thread-local last_error accessor — intact, upstream touched ggml-rpc.cpp by one unrelated line. * 3679b23 Fixes (RPC) — intact. * 2aa76c7 fix(metal): drop stray kernel_pad_f32 — still applies; upstream has since refactored pad into a templated kernel_pad_impl<T>, so the duplicate definition that referenced the nonexistent kargs_pad.s0..s3 is gone on both sides and nothing had to be re-applied.

* model: add MTP support for Nemotron Nano model * model: add mtp_flags for nemotron model * address review comments

* model: add MTP support for Nemotron Nano model * model: add mtp_flags for nemotron model * address review comments (cherry picked from commit 7a20b41)

* model: add MTP support for Nemotron Nano model * model: add mtp_flags for nemotron model * address review comments

ggml-org#26725 added MTP export to NemotronHModel, keyed on num_nextn_predict_layers. Puzzle's config carries that key, but NemotronHPuzzleModel bypasses NemotronHModel.__init__ (its per-block config needs a different setup), so _mtp_bid was never assigned and modify_tensors raised AttributeError on any mtp.* tensor. Puzzle's head is also laid out by mtp_block_configs, not the mtp.layers.* form the base maps. Set _mtp_bid to None, drop mtp.* in filter_tensors, and declare supports_mtp_export = False so --mtp / --no-mtp fail at the CLI.

…#25444) * hparams: add per-layer n_ff_exp/n_expert_used arrays with scalar-or-array loading G1/G2 infrastructure for variable-per-layer expert FFN size and top-k routing (required for Puzzle-75B which has 5 distinct n_ff_exp values and 7 top-k values across its 40 MoE layers). Design: rename scalar members to _impl suffix (following existing convention), add LLAMA_MAX_LAYERS arrays, add n_ff_exp(il)/n_expert_used(il) accessors with scalar fallback. No new GGUF keys: reuses existing expert_feed_forward_length and expert_used_count keys via get_key_or_arr (scalar -> broadcast, array -> per-layer). - llama-hparams.h: n_ff_exp -> n_ff_exp_impl, n_expert_used -> n_expert_used_impl; add n_ff_exp_arr / n_expert_used_arr arrays; add per-layer accessor declarations. - llama-hparams.cpp: implement n_ff_exp(il) and n_expert_used(il); out-of-range il returns impl safely (shared code, no abort). - llama-model.cpp: central n_expert_used load changed to get_key_or_arr; derive impl as max-of-array for validations and backward compat; zero both new arrays; HunyuanVL override also zeroes n_expert_used_arr. - llama-graph.cpp: aggregation loop in build_moe_ffn uses hparams.n_expert_used(il) so per-layer top-k bounds the ggml_view loop correctly. - All other files: mechanical rename hparams.n_{ff_exp,expert_used} -> *_impl. Scalar arches are unaffected (broadcast fills all array slots with the single value). (cherry picked from commit 269a81e) * nemotron-h: use per-layer n_ff_exp(il) and n_expert_used(il) at MoE call-sites Load n_ff_exp via get_key_or_arr into hparams.n_ff_exp_arr in load_arch_hparams; derive impl as max for existing uniform GGUFs. In load_arch_tensors, compute n_ff_exp_i = hparams.n_ff_exp(i) with fallback to n_ff(i)/n_expert_used(i) for GGUFs that omit expert_feed_forward_length. In build_ffn_layer, pass hparams.n_expert_used(il) to build_moe_ffn so per-layer top-k is used for expert routing selection. All other nemotron-h behaviour (mamba2, attention, shared-exp, latent projection, routed_scaling_factor, expert_weights_norm, sigmoid gating) is unchanged. (cherry picked from commit b1878a1) * arch/*.cpp + gguf-py: mechanical rename n_ff_exp->n_ff_exp_impl, n_expert_used->n_expert_used_impl All non-nemotron arch files continue using the scalar impl member directly. Behaviour is identical: the impl value is the broadcast value from the GGUF scalar. gguf_writer: add_expert_feed_forward_length and add_expert_used_count now accept int | Sequence[int], mirroring add_feed_forward_length, so converters can write per-layer arrays with the same existing GGUF keys. (cherry picked from commit 8f009f5) * convert: support NemotronHPuzzleForCausalLM (per-block MoE config) Parse block_configs/mtp_block_configs into per-layer arrays (scalar-or-array keys), append the MTP [attention, moe] sub-blocks as blk.88/blk.89 with nextn tensors, accept the backbone.* prefix, and register the arch. Also fix a pre-existing undeclared _experts attribute on NemotronHModel. (cherry picked from commit d1a592f) * nemotron-h: distinguish Nemotron 3 Puzzle (75B.A9B) from Super (120B.A12B) Both have 88 layers; the per-layer expert_used_count array (heterogeneous for Puzzle, broadcast-uniform for Super) is the discriminator. (cherry picked from commit f824e09) * convert: accept the official Puzzle BF16 checkpoint's tensor naming The officially distributed BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle- 75B-A9B-BF16) names the trunk model.* (model.layers.*, model.embeddings, model.norm_f) where the original release used the NemotronH-style backbone.*, and spells the router bias e_score_correction_bias instead of e_score_correction.bias. Normalize both at the top of NemotronHPuzzleModel.modify_tensors so either checkpoint converts; every tensor name in the official index (42683 keys, MTP head included) resolves through the tensor map after normalization. (cherry picked from commit 189b67f) * laguna: use n_ff_exp_impl for the uniform-MoE FFN size Laguna landed after this branch was cut and reads hparams.n_ff_exp as a scalar. This series turns it into a per-layer array with an n_ff_exp(il) accessor, so the three scalar reads no longer compile. Laguna is a uniform MoE, so point them at the scalar fallback n_ff_exp_impl, same as deepseek2/qwen3moe/gemma4 in this series. No behaviour change. (cherry picked from commit dbedc9e) * arch: extend the n_ff_exp/n_expert_used rename to archs added upstream kimi-k3, dflash, bailingmoe3, deepseek4, granite-swa and the nemotron-h MTP block still referenced the scalar fields by their old names. n_ff_exp and n_expert_used are accessors now, so those reads no longer compile; point the non-per-layer archs at the _impl scalars and use the indexed form where the call site is per-layer. * convert: keep Puzzle opted out of the NemotronH MTP export path #26725 added MTP export to NemotronHModel, keyed on num_nextn_predict_layers. Puzzle's config carries that key, but NemotronHPuzzleModel bypasses NemotronHModel.__init__ (its per-block config needs a different setup), so _mtp_bid was never assigned and modify_tensors raised AttributeError on any mtp.* tensor. Puzzle's head is also laid out by mtp_block_configs, not the mtp.layers.* form the base maps. Set _mtp_bid to None, drop mtp.* in filter_tensors, and declare supports_mtp_export = False so --mtp / --no-mtp fail at the CLI. * llama: replace n_ff_exp/n_expert_used scalars with per-layer accessors Follow-up to review feedback: the previous revision kept the scalar hparams fields alongside the new per-layer arrays, which duplicated state that get_key_or_arr already handles by broadcasting a scalar value over every layer. Drop both scalars and expose n_ff_exp(il) / n_expert_used(il) built exactly like the existing n_head_kv(il) and n_ff(il) accessors: they index the array and GGML_ABORT out of range, with il defaulting to 0 so genuinely uniform call sites stay a plain n_ff_exp(). Arch loaders now read both keys through get_key_or_arr over n_layer_all, and the n_expert_used validation checks the maximum across layers instead of a single field. * llama: restore per-key required flags on the expert hparam reads The scalar-to-array conversion passed required=false at every call site, which silently made mandatory keys optional. Each read now carries the same required flag it had before the conversion.

…ggml-org#25444) * hparams: add per-layer n_ff_exp/n_expert_used arrays with scalar-or-array loading G1/G2 infrastructure for variable-per-layer expert FFN size and top-k routing (required for Puzzle-75B which has 5 distinct n_ff_exp values and 7 top-k values across its 40 MoE layers). Design: rename scalar members to _impl suffix (following existing convention), add LLAMA_MAX_LAYERS arrays, add n_ff_exp(il)/n_expert_used(il) accessors with scalar fallback. No new GGUF keys: reuses existing expert_feed_forward_length and expert_used_count keys via get_key_or_arr (scalar -> broadcast, array -> per-layer). - llama-hparams.h: n_ff_exp -> n_ff_exp_impl, n_expert_used -> n_expert_used_impl; add n_ff_exp_arr / n_expert_used_arr arrays; add per-layer accessor declarations. - llama-hparams.cpp: implement n_ff_exp(il) and n_expert_used(il); out-of-range il returns impl safely (shared code, no abort). - llama-model.cpp: central n_expert_used load changed to get_key_or_arr; derive impl as max-of-array for validations and backward compat; zero both new arrays; HunyuanVL override also zeroes n_expert_used_arr. - llama-graph.cpp: aggregation loop in build_moe_ffn uses hparams.n_expert_used(il) so per-layer top-k bounds the ggml_view loop correctly. - All other files: mechanical rename hparams.n_{ff_exp,expert_used} -> *_impl. Scalar arches are unaffected (broadcast fills all array slots with the single value). (cherry picked from commit 269a81e) * nemotron-h: use per-layer n_ff_exp(il) and n_expert_used(il) at MoE call-sites Load n_ff_exp via get_key_or_arr into hparams.n_ff_exp_arr in load_arch_hparams; derive impl as max for existing uniform GGUFs. In load_arch_tensors, compute n_ff_exp_i = hparams.n_ff_exp(i) with fallback to n_ff(i)/n_expert_used(i) for GGUFs that omit expert_feed_forward_length. In build_ffn_layer, pass hparams.n_expert_used(il) to build_moe_ffn so per-layer top-k is used for expert routing selection. All other nemotron-h behaviour (mamba2, attention, shared-exp, latent projection, routed_scaling_factor, expert_weights_norm, sigmoid gating) is unchanged. (cherry picked from commit b1878a1) * arch/*.cpp + gguf-py: mechanical rename n_ff_exp->n_ff_exp_impl, n_expert_used->n_expert_used_impl All non-nemotron arch files continue using the scalar impl member directly. Behaviour is identical: the impl value is the broadcast value from the GGUF scalar. gguf_writer: add_expert_feed_forward_length and add_expert_used_count now accept int | Sequence[int], mirroring add_feed_forward_length, so converters can write per-layer arrays with the same existing GGUF keys. (cherry picked from commit 8f009f5) * convert: support NemotronHPuzzleForCausalLM (per-block MoE config) Parse block_configs/mtp_block_configs into per-layer arrays (scalar-or-array keys), append the MTP [attention, moe] sub-blocks as blk.88/blk.89 with nextn tensors, accept the backbone.* prefix, and register the arch. Also fix a pre-existing undeclared _experts attribute on NemotronHModel. (cherry picked from commit d1a592f) * nemotron-h: distinguish Nemotron 3 Puzzle (75B.A9B) from Super (120B.A12B) Both have 88 layers; the per-layer expert_used_count array (heterogeneous for Puzzle, broadcast-uniform for Super) is the discriminator. (cherry picked from commit f824e09) * convert: accept the official Puzzle BF16 checkpoint's tensor naming The officially distributed BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle- 75B-A9B-BF16) names the trunk model.* (model.layers.*, model.embeddings, model.norm_f) where the original release used the NemotronH-style backbone.*, and spells the router bias e_score_correction_bias instead of e_score_correction.bias. Normalize both at the top of NemotronHPuzzleModel.modify_tensors so either checkpoint converts; every tensor name in the official index (42683 keys, MTP head included) resolves through the tensor map after normalization. (cherry picked from commit 189b67f) * laguna: use n_ff_exp_impl for the uniform-MoE FFN size Laguna landed after this branch was cut and reads hparams.n_ff_exp as a scalar. This series turns it into a per-layer array with an n_ff_exp(il) accessor, so the three scalar reads no longer compile. Laguna is a uniform MoE, so point them at the scalar fallback n_ff_exp_impl, same as deepseek2/qwen3moe/gemma4 in this series. No behaviour change. (cherry picked from commit dbedc9e) * arch: extend the n_ff_exp/n_expert_used rename to archs added upstream kimi-k3, dflash, bailingmoe3, deepseek4, granite-swa and the nemotron-h MTP block still referenced the scalar fields by their old names. n_ff_exp and n_expert_used are accessors now, so those reads no longer compile; point the non-per-layer archs at the _impl scalars and use the indexed form where the call site is per-layer. * convert: keep Puzzle opted out of the NemotronH MTP export path ggml-org#26725 added MTP export to NemotronHModel, keyed on num_nextn_predict_layers. Puzzle's config carries that key, but NemotronHPuzzleModel bypasses NemotronHModel.__init__ (its per-block config needs a different setup), so _mtp_bid was never assigned and modify_tensors raised AttributeError on any mtp.* tensor. Puzzle's head is also laid out by mtp_block_configs, not the mtp.layers.* form the base maps. Set _mtp_bid to None, drop mtp.* in filter_tensors, and declare supports_mtp_export = False so --mtp / --no-mtp fail at the CLI. * llama: replace n_ff_exp/n_expert_used scalars with per-layer accessors Follow-up to review feedback: the previous revision kept the scalar hparams fields alongside the new per-layer arrays, which duplicated state that get_key_or_arr already handles by broadcasting a scalar value over every layer. Drop both scalars and expose n_ff_exp(il) / n_expert_used(il) built exactly like the existing n_head_kv(il) and n_ff(il) accessors: they index the array and GGML_ABORT out of range, with il defaulting to 0 so genuinely uniform call sites stay a plain n_ff_exp(). Arch loaders now read both keys through get_key_or_arr over n_layer_all, and the n_expert_used validation checks the maximum across layers instead of a single field. * llama: restore per-key required flags on the expert hparam reads The scalar-to-array conversion passed required=false at every call site, which silently made mandatory keys optional. Each read now carries the same required flag it had before the conversion.

* model: add MTP support for Nemotron Nano model * model: add mtp_flags for nemotron model * address review comments

…ggml-org#25444) * hparams: add per-layer n_ff_exp/n_expert_used arrays with scalar-or-array loading G1/G2 infrastructure for variable-per-layer expert FFN size and top-k routing (required for Puzzle-75B which has 5 distinct n_ff_exp values and 7 top-k values across its 40 MoE layers). Design: rename scalar members to _impl suffix (following existing convention), add LLAMA_MAX_LAYERS arrays, add n_ff_exp(il)/n_expert_used(il) accessors with scalar fallback. No new GGUF keys: reuses existing expert_feed_forward_length and expert_used_count keys via get_key_or_arr (scalar -> broadcast, array -> per-layer). - llama-hparams.h: n_ff_exp -> n_ff_exp_impl, n_expert_used -> n_expert_used_impl; add n_ff_exp_arr / n_expert_used_arr arrays; add per-layer accessor declarations. - llama-hparams.cpp: implement n_ff_exp(il) and n_expert_used(il); out-of-range il returns impl safely (shared code, no abort). - llama-model.cpp: central n_expert_used load changed to get_key_or_arr; derive impl as max-of-array for validations and backward compat; zero both new arrays; HunyuanVL override also zeroes n_expert_used_arr. - llama-graph.cpp: aggregation loop in build_moe_ffn uses hparams.n_expert_used(il) so per-layer top-k bounds the ggml_view loop correctly. - All other files: mechanical rename hparams.n_{ff_exp,expert_used} -> *_impl. Scalar arches are unaffected (broadcast fills all array slots with the single value). (cherry picked from commit 269a81e) * nemotron-h: use per-layer n_ff_exp(il) and n_expert_used(il) at MoE call-sites Load n_ff_exp via get_key_or_arr into hparams.n_ff_exp_arr in load_arch_hparams; derive impl as max for existing uniform GGUFs. In load_arch_tensors, compute n_ff_exp_i = hparams.n_ff_exp(i) with fallback to n_ff(i)/n_expert_used(i) for GGUFs that omit expert_feed_forward_length. In build_ffn_layer, pass hparams.n_expert_used(il) to build_moe_ffn so per-layer top-k is used for expert routing selection. All other nemotron-h behaviour (mamba2, attention, shared-exp, latent projection, routed_scaling_factor, expert_weights_norm, sigmoid gating) is unchanged. (cherry picked from commit b1878a1) * arch/*.cpp + gguf-py: mechanical rename n_ff_exp->n_ff_exp_impl, n_expert_used->n_expert_used_impl All non-nemotron arch files continue using the scalar impl member directly. Behaviour is identical: the impl value is the broadcast value from the GGUF scalar. gguf_writer: add_expert_feed_forward_length and add_expert_used_count now accept int | Sequence[int], mirroring add_feed_forward_length, so converters can write per-layer arrays with the same existing GGUF keys. (cherry picked from commit 8f009f5) * convert: support NemotronHPuzzleForCausalLM (per-block MoE config) Parse block_configs/mtp_block_configs into per-layer arrays (scalar-or-array keys), append the MTP [attention, moe] sub-blocks as blk.88/blk.89 with nextn tensors, accept the backbone.* prefix, and register the arch. Also fix a pre-existing undeclared _experts attribute on NemotronHModel. (cherry picked from commit d1a592f) * nemotron-h: distinguish Nemotron 3 Puzzle (75B.A9B) from Super (120B.A12B) Both have 88 layers; the per-layer expert_used_count array (heterogeneous for Puzzle, broadcast-uniform for Super) is the discriminator. (cherry picked from commit f824e09) * convert: accept the official Puzzle BF16 checkpoint's tensor naming The officially distributed BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle- 75B-A9B-BF16) names the trunk model.* (model.layers.*, model.embeddings, model.norm_f) where the original release used the NemotronH-style backbone.*, and spells the router bias e_score_correction_bias instead of e_score_correction.bias. Normalize both at the top of NemotronHPuzzleModel.modify_tensors so either checkpoint converts; every tensor name in the official index (42683 keys, MTP head included) resolves through the tensor map after normalization. (cherry picked from commit 189b67f) * laguna: use n_ff_exp_impl for the uniform-MoE FFN size Laguna landed after this branch was cut and reads hparams.n_ff_exp as a scalar. This series turns it into a per-layer array with an n_ff_exp(il) accessor, so the three scalar reads no longer compile. Laguna is a uniform MoE, so point them at the scalar fallback n_ff_exp_impl, same as deepseek2/qwen3moe/gemma4 in this series. No behaviour change. (cherry picked from commit dbedc9e) * arch: extend the n_ff_exp/n_expert_used rename to archs added upstream kimi-k3, dflash, bailingmoe3, deepseek4, granite-swa and the nemotron-h MTP block still referenced the scalar fields by their old names. n_ff_exp and n_expert_used are accessors now, so those reads no longer compile; point the non-per-layer archs at the _impl scalars and use the indexed form where the call site is per-layer. * convert: keep Puzzle opted out of the NemotronH MTP export path ggml-org#26725 added MTP export to NemotronHModel, keyed on num_nextn_predict_layers. Puzzle's config carries that key, but NemotronHPuzzleModel bypasses NemotronHModel.__init__ (its per-block config needs a different setup), so _mtp_bid was never assigned and modify_tensors raised AttributeError on any mtp.* tensor. Puzzle's head is also laid out by mtp_block_configs, not the mtp.layers.* form the base maps. Set _mtp_bid to None, drop mtp.* in filter_tensors, and declare supports_mtp_export = False so --mtp / --no-mtp fail at the CLI. * llama: replace n_ff_exp/n_expert_used scalars with per-layer accessors Follow-up to review feedback: the previous revision kept the scalar hparams fields alongside the new per-layer arrays, which duplicated state that get_key_or_arr already handles by broadcasting a scalar value over every layer. Drop both scalars and expose n_ff_exp(il) / n_expert_used(il) built exactly like the existing n_head_kv(il) and n_ff(il) accessors: they index the array and GGML_ABORT out of range, with il defaulting to 0 so genuinely uniform call sites stay a plain n_ff_exp(). Arch loaders now read both keys through get_key_or_arr over n_layer_all, and the n_expert_used validation checks the maximum across layers instead of a single field. * llama: restore per-key required flags on the expert hparam reads The scalar-to-array conversion passed required=false at every call site, which silently made mandatory keys optional. Each read now carries the same required flag it had before the conversion.

* model: add MTP support for Nemotron Nano model * model: add mtp_flags for nemotron model * address review comments

…ggml-org#25444) * hparams: add per-layer n_ff_exp/n_expert_used arrays with scalar-or-array loading G1/G2 infrastructure for variable-per-layer expert FFN size and top-k routing (required for Puzzle-75B which has 5 distinct n_ff_exp values and 7 top-k values across its 40 MoE layers). Design: rename scalar members to _impl suffix (following existing convention), add LLAMA_MAX_LAYERS arrays, add n_ff_exp(il)/n_expert_used(il) accessors with scalar fallback. No new GGUF keys: reuses existing expert_feed_forward_length and expert_used_count keys via get_key_or_arr (scalar -> broadcast, array -> per-layer). - llama-hparams.h: n_ff_exp -> n_ff_exp_impl, n_expert_used -> n_expert_used_impl; add n_ff_exp_arr / n_expert_used_arr arrays; add per-layer accessor declarations. - llama-hparams.cpp: implement n_ff_exp(il) and n_expert_used(il); out-of-range il returns impl safely (shared code, no abort). - llama-model.cpp: central n_expert_used load changed to get_key_or_arr; derive impl as max-of-array for validations and backward compat; zero both new arrays; HunyuanVL override also zeroes n_expert_used_arr. - llama-graph.cpp: aggregation loop in build_moe_ffn uses hparams.n_expert_used(il) so per-layer top-k bounds the ggml_view loop correctly. - All other files: mechanical rename hparams.n_{ff_exp,expert_used} -> *_impl. Scalar arches are unaffected (broadcast fills all array slots with the single value). (cherry picked from commit 269a81e) * nemotron-h: use per-layer n_ff_exp(il) and n_expert_used(il) at MoE call-sites Load n_ff_exp via get_key_or_arr into hparams.n_ff_exp_arr in load_arch_hparams; derive impl as max for existing uniform GGUFs. In load_arch_tensors, compute n_ff_exp_i = hparams.n_ff_exp(i) with fallback to n_ff(i)/n_expert_used(i) for GGUFs that omit expert_feed_forward_length. In build_ffn_layer, pass hparams.n_expert_used(il) to build_moe_ffn so per-layer top-k is used for expert routing selection. All other nemotron-h behaviour (mamba2, attention, shared-exp, latent projection, routed_scaling_factor, expert_weights_norm, sigmoid gating) is unchanged. (cherry picked from commit b1878a1) * arch/*.cpp + gguf-py: mechanical rename n_ff_exp->n_ff_exp_impl, n_expert_used->n_expert_used_impl All non-nemotron arch files continue using the scalar impl member directly. Behaviour is identical: the impl value is the broadcast value from the GGUF scalar. gguf_writer: add_expert_feed_forward_length and add_expert_used_count now accept int | Sequence[int], mirroring add_feed_forward_length, so converters can write per-layer arrays with the same existing GGUF keys. (cherry picked from commit 8f009f5) * convert: support NemotronHPuzzleForCausalLM (per-block MoE config) Parse block_configs/mtp_block_configs into per-layer arrays (scalar-or-array keys), append the MTP [attention, moe] sub-blocks as blk.88/blk.89 with nextn tensors, accept the backbone.* prefix, and register the arch. Also fix a pre-existing undeclared _experts attribute on NemotronHModel. (cherry picked from commit d1a592f) * nemotron-h: distinguish Nemotron 3 Puzzle (75B.A9B) from Super (120B.A12B) Both have 88 layers; the per-layer expert_used_count array (heterogeneous for Puzzle, broadcast-uniform for Super) is the discriminator. (cherry picked from commit f824e09) * convert: accept the official Puzzle BF16 checkpoint's tensor naming The officially distributed BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle- 75B-A9B-BF16) names the trunk model.* (model.layers.*, model.embeddings, model.norm_f) where the original release used the NemotronH-style backbone.*, and spells the router bias e_score_correction_bias instead of e_score_correction.bias. Normalize both at the top of NemotronHPuzzleModel.modify_tensors so either checkpoint converts; every tensor name in the official index (42683 keys, MTP head included) resolves through the tensor map after normalization. (cherry picked from commit 189b67f) * laguna: use n_ff_exp_impl for the uniform-MoE FFN size Laguna landed after this branch was cut and reads hparams.n_ff_exp as a scalar. This series turns it into a per-layer array with an n_ff_exp(il) accessor, so the three scalar reads no longer compile. Laguna is a uniform MoE, so point them at the scalar fallback n_ff_exp_impl, same as deepseek2/qwen3moe/gemma4 in this series. No behaviour change. (cherry picked from commit dbedc9e) * arch: extend the n_ff_exp/n_expert_used rename to archs added upstream kimi-k3, dflash, bailingmoe3, deepseek4, granite-swa and the nemotron-h MTP block still referenced the scalar fields by their old names. n_ff_exp and n_expert_used are accessors now, so those reads no longer compile; point the non-per-layer archs at the _impl scalars and use the indexed form where the call site is per-layer. * convert: keep Puzzle opted out of the NemotronH MTP export path ggml-org#26725 added MTP export to NemotronHModel, keyed on num_nextn_predict_layers. Puzzle's config carries that key, but NemotronHPuzzleModel bypasses NemotronHModel.__init__ (its per-block config needs a different setup), so _mtp_bid was never assigned and modify_tensors raised AttributeError on any mtp.* tensor. Puzzle's head is also laid out by mtp_block_configs, not the mtp.layers.* form the base maps. Set _mtp_bid to None, drop mtp.* in filter_tensors, and declare supports_mtp_export = False so --mtp / --no-mtp fail at the CLI. * llama: replace n_ff_exp/n_expert_used scalars with per-layer accessors Follow-up to review feedback: the previous revision kept the scalar hparams fields alongside the new per-layer arrays, which duplicated state that get_key_or_arr already handles by broadcasting a scalar value over every layer. Drop both scalars and expose n_ff_exp(il) / n_expert_used(il) built exactly like the existing n_head_kv(il) and n_ff(il) accessors: they index the array and GGML_ABORT out of range, with il defaulting to 0 so genuinely uniform call sites stay a plain n_ff_exp(). Arch loaders now read both keys through get_key_or_arr over n_layer_all, and the n_expert_used validation checks the maximum across layers instead of a single field. * llama: restore per-key required flags on the expert hparam reads The scalar-to-array conversion passed required=false at every call site, which silently made mandatory keys optional. Each read now carries the same required flag it had before the conversion.

model: add MTP support for Nemotron Nano model

2f778a3

ruixiang63 requested a review from ggerganov August 7, 2026 13:46

github-actions Bot added model Model specific conversion labels Aug 7, 2026

ruixiang63 requested a review from am17an August 7, 2026 13:47

am17an approved these changes Aug 7, 2026

View reviewed changes

ggerganov marked this pull request as ready for review August 7, 2026 18:46

ggerganov requested a review from CISC as a code owner August 7, 2026 18:46

CISC reviewed Aug 7, 2026

View reviewed changes

Comment thread src/models/nemotron-h.cpp

gaugarg-nv reviewed Aug 8, 2026

View reviewed changes

Comment thread src/models/nemotron-h-moe.cpp Outdated

gaugarg-nv reviewed Aug 8, 2026

View reviewed changes

ruixiang63 added 2 commits August 10, 2026 00:57

model: add mtp_flags for nemotron model

00d29e1

address review comments

6c57877

CISC approved these changes Aug 10, 2026

View reviewed changes

ggerganov merged commit 7a20b41 into ggml-org:master Aug 10, 2026
25 of 29 checks passed

ruixiang63 deleted the nemotron-nano-mtp branch August 10, 2026 20:27

ynankani mentioned this pull request Aug 11, 2026

Keep lm-head quantization scales for nemotron --mtp export #26903

Merged

crusaderky pushed a commit to crusaderky/llama.cpp that referenced this pull request Aug 12, 2026

model: add MTP support for Nemotron model (ggml-org#26725)

df44916

* model: add MTP support for Nemotron Nano model * model: add mtp_flags for nemotron model * address review comments

huaxel pushed a commit to huaxel/UmaLlama that referenced this pull request Aug 12, 2026

model: add MTP support for Nemotron model (ggml-org#26725)

21d91e1

* model: add MTP support for Nemotron Nano model * model: add mtp_flags for nemotron model * address review comments

Piggidragon mentioned this pull request Aug 13, 2026

MTP bug: NVIDIA-Nemotron-3.5-Lightning crash TheTom/llama-cpp-turboquant#296

Open

brittlewis12 pushed a commit to brittlewis12/llama.cpp that referenced this pull request Aug 17, 2026

model: add MTP support for Nemotron model (ggml-org#26725)

ff32606

* model: add MTP support for Nemotron Nano model * model: add mtp_flags for nemotron model * address review comments

YanissAmz mentioned this pull request Aug 23, 2026

model: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support #25444

Merged

thecodacus pushed a commit to thecodacus/llama.cpp that referenced this pull request Sep 7, 2026

model: add MTP support for Nemotron model (ggml-org#26725)

57d0112

* model: add MTP support for Nemotron Nano model * model: add mtp_flags for nemotron model * address review comments

zbrad pushed a commit to zbrad/llama.cpp that referenced this pull request Sep 10, 2026

model: add MTP support for Nemotron model (ggml-org#26725)

4c1b6a4

* model: add MTP support for Nemotron Nano model * model: add mtp_flags for nemotron model * address review comments

frozenblade1224 mentioned this pull request Sep 17, 2026

model : extend Nemotron MTP support #29018

Merged

ruixiang63 mentioned this pull request Sep 23, 2026

models: Qwen3.8-Flash-Next MTP #28243

Closed

christopherthompson81 mentioned this pull request Oct 22, 2026

Nemotron-H: support the hybrid Mamba-2 text models, their MTP head, and the VL / Omni wrappers christopherthompson81/inference.rs#565

Open

Navigation Menu

Repository navigation

model: add MTP support for Nemotron model - #26725

model: add MTP support for Nemotron model#26725
ggerganov merged 3 commits into
ggml-org:masterfrom
ruixiang63:nemotron-nano-mtp

ruixiang63 commented Aug 7, 2026 •

edited

Loading

Uh oh!

am17an left a comment

Uh oh!

Uh oh!

Uh oh!

gaugarg-nv Aug 8, 2026

Uh oh!

ruixiang63 Aug 9, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

ruixiang63 commented Aug 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Overview

Additional information

Requirements

Uh oh!

am17an left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

gaugarg-nv Aug 8, 2026

Choose a reason for hiding this comment

Uh oh!

ruixiang63 Aug 9, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

ruixiang63 commented Aug 7, 2026 •

edited

Loading