* initial map to load deepseek 4 arch
* wip
* wip: match graph build and attn logic for dpv4
* wip: Enhance DeepSeek-V4 architecture with new tensor types and sqrtsoftplus gating function
* Update DeepSeek-V4 to support raw key indexing with read/write indices
* fix mismatch in attn_raw
* Enable FA with CSA/HCA
* Fix logit mismatch with FA path
* Clean traces and logs for debug
* Refactor DSV4 tensor handling for MTP execution and improve raw context management
* Refactor DeepSeek4 tensor operations: replace manual weighted sum and post-processing with new helper functions
* Share mHC pre-projection and fix packed DSV4 writes
* DSV4: add shared top-k selection and improve mask handling
* Fix DSV4 c2048 view stride and duplicate loader instantiation
* Reuse shared RMS normalization in DSV4 graph
* Replace DSV4 indexer rotation with shared Hadamard
* Share CSA visibility mask with DSV4 LID
* dsv4: document dependency ordering and reset state
* Remove DSV4 zero-dependency graph shim
* Fix DSV4 packed stream execution
* Remove DSV4 l_out backend override
* Enable DSV4 quantized K-only cache
* Revert "Enable DSV4 quantized K-only cache"
This reverts commit 04f9b425321f62ba60e16d1bea2f8de714cfe855.
* Fix DSV4 quantized cache accounting
* Fail closed on unsupported DSV4 cache lifecycle operations
* Various optimizations
* llama: fix GGML_METAL=ON build - missing ggml-metal.h include in llama-dflash.cpp (#2134)
llama-dflash.cpp calls ggml_backend_is_metal() and
ggml_backend_metal_set_n_cb() inside an #ifdef GGML_USE_METAL block but
never includes ggml-metal.h, so any Metal-enabled build fails to
compile. Add the same guarded include llama.cpp already uses.
* New op: ggml_sum_rows_ext (#2132)
* Add ggml_sum_rows_ext
* openPangu: use ggml_sum_rows_ext also in mhc_post
* openPangu: use ggml_sum_rows_ext also in mhc_tail
* Minor
* Reuse shared inverse RoPE operation for DSV4
* Reuse maintainer CUDA concat implementation
* WIP
* hc_pre
* hc_post
* Remove unnecessary mask manipulations
* WIP
* Take into account swiglu limits
* Turn on fused indexer by default
* Give names to mat mul results
* More named ops
* dsv4: do not uselessly copy the KV cache
+20% TG at 32k tokens
* mask_to_index and make CPU FA work with that
* Much better CPU-only, CUDA still not functional
* Better CPU TG
I'm now at 9.7 t/s for zero context and 6.5 t/s for context of 32k.
PP is 120 t/s for short context and 101 t/s at 32k.
* Even better CPU TG
I'm now at 8.1 t/s for context of 32k tokens.
* Turn off DSA on CUDA for now
* Fix CUDA DSA
* Remove again the unnecessary softmax result buffer
* Experiments
* Various
* More named ops
* Forgot to uncomment
---------
Co-authored-by: samuel <samueloliveira32df@gmail.com>
Co-authored-by: hchengit <95317477+hchengit@users.noreply.github.com>
* openpangu: Stage-1 converter probe for openPangu-2.0-Flash
Add OpenPanguV2ForCausalLM conversion support (converter-only; runtime graph
is Stage-2). Registers a new LLM_ARCH_OPENPANGU on the Python/gguf-py side:
- gguf-py/constants.py: MODEL_ARCH.OPENPANGU + name, indexer KV keys, 22 new
tensor enums (DSA indexer x4, MoME convs x3, param-sink x2, mHC/Hyper-
Connections x12, block-post-norm), and the full MODEL_TENSORS list reusing
the deepseek MLA + MoE + NextN bricks.
- tensor_mapping.py: arch-specific block mappings that disambiguate the
sandwich norms (post_attention/pre_mlp/post_mlp) and pin every Pangu-only
tensor; non-block global mHC merge module.
- convert_hf_to_gguf.py: OpenPanguV2Model (subclasses DeepseekV2Model) with
set_gguf_parameters (MLA/MoE/indexer/mHC/param-sink/DSA+SWA metadata),
modify_tensors (expert merge, kv_b split, no MTP skip), and the
OpenPanguV2Tokenizer pre-tokenizer hash.
Validated offline against the real 50-shard safetensors index: all 37,587
tensors map to a GGUF target (0 unmapped), and set_gguf_parameters reads only
hparams present in config.json. No weights downloaded; no GPU. Pinned on the
ik/dsa_loop_hadamard_blend DSA substrate.
* openpangu: Stage-2 arch scaffold (LLM_ARCH_OPENPANGU) — loadable, compiles
New arch on main (DSA-decoupled). Declares openPangu-2.0-Flash to the runtime so
the model loads into memory; the compute graph is the next step.
- llama-arch.{h,cpp}: LLM_ARCH_OPENPANGU + name; 3 KV keys (mhc_num_stream,
mhc_recur_norm, param_sink_number); 18 tensor enums (mHC x12, MoME conv x3,
param-sink x2, block-post-norm).
- llama-model.cpp: OPENPANGU tensor-name block, strings matched to the converter.
- llama-model.h: layer + model struct fields (mHC / conv / sink / block-post / merge).
- llama-hparams.{h,cpp}: reader (MLA + MoE + sigmoid gate + indexer + mHC +
param-sink + NextN); n_layer_kv_from_start = n_layer - nextn (MTP skipped).
- llama-load-tensors.cpp: create_openpangu_tensors (GLM-DSA MLA/MoE base + Pangu
tensors; indexer loaded-but-unused for dense fallback); dispatch + is_mla_attn.
Builds clean (CPU-only libllama). Dense-fallback design: no DSA indexer / SWA
windowing / MTP for first generation (exact <=512 tokens). Graph is Stage-2b.
* openpangu: fix compresskv_conv dim (kv_lora_rank, not +rope); pin attention order in spec
* openpangu: end-to-end runtime — build_openpangu graph runs, generates (garbled)
First full forward pass of openPangu-2.0-Flash on ik_llama. Pipeline works end to
end: new LLM_ARCH_OPENPANGU loads the Q4 GGUF, the graph executes, and llama-cli
generates 40 tokens (EXIT=0). Output is currently garbled (tensor-layout bug to
debug), but the structure is proven.
graphs/build_openpangu.cpp: dense decompressed-MHA attention + 4-stream mHC
(Hyper-Connections) with 20-iter Sinkhorn + MoE(sigmoid+shared) + sandwich norms
+ entry stream-repeat/tail-merge + inp_out_ids selection.
Bring-up fixes to load+run:
- llama-vocab.cpp: register 'openpangu' pre-tokenizer (QWEN2 family)
- llama.cpp: OPENPANGU -> LLAMA_ROPE_TYPE_NORM (was defaulting to NONE=-1)
- llama-load-tensors.cpp: full wkv_b load; k_b/v_b as flattened 2D; block_post_norm
dim = S*H (10240); conv weights 2D {3,C}; mHC alpha/beta/gamma + param_sink +
merge params use bare (no-.weight) tensor names
- llama-model.cpp: OPENPANGU is NOT is_mla_attn (decompressed MHA, standard KV cache)
- graph loop bounded to base layers (skip NextN/MTP)
v0 deferrals (need conv-state cache / manual attention path, all documented):
MoME convs (passthrough), o_conv, param_sink. Next: fix the layout bug to coherence.
* openpangu: COHERENT generation — NEOX rope, Sinkhorn orientation, MoME convs, param_sink
Four correctness fixes on top of the end-to-end scaffold, verified checkpoint-by-
checkpoint against a Python golden reference running on the GGUF's own dequantized
weights (block-0 activations now match to rounding at full fidelity):
- rope: NORM -> NEOX. Pangu config rope_interleave=false; the Infer source maps it
as is_neox_style = not rope_interleave (rotary_mode='half').
- mHC Sinkhorn: the flat h_res block is torch-[r,c] row-major, so a bare ggml
reshape lands column-fastest; the doubly-stochastic iteration ran transposed
(Sinkhorn is not transpose-symmetric). One transpose at input fixes the whole
chain including the mhc_post application.
- MoME convs (qa/compresskv/o): were passthrough stubs. Implemented as
out = x + causal_conv1d(x) (every Infer call site uses residual_connection=1;
tap stats confirm the perturbation form). Taps cast f16->f32 for ggml_mul.
Batch-local v0: exact for fresh-sequence prefill; decode steps miss the
t-1/t-2 taps until a conv-state cache exists.
- param_sink: 128 learned latent-KV entries prepended per layer via a manual
attention path (kv_store + explicit soft_max over [sinks ++ cache]); huge
effect at short context. o_conv now applied pre-o_proj on the same path.
flash_attn forced off for OPENPANGU (FA kernel cannot see the sinks).
- converter: add_bos_token=true (HF prepends <|pangu_text_start|> via the
post-processor; the key was absent so ik dropped BOS).
Greedy Q4_K_M smoke, chat template + <think>: coherent CoT reasoning and a
correct answer. Layer-0 instrumentation (opg0_* names) kept for now.
* openpangu: MoME conv-state cache — decode steps get real t-1/t-2 taps
Allocate a per-layer cache_s_l tensor for OPENPANGU base layers holding the last
two pre-conv latents of the three MoME sites, packed
[qa 2*1024 | compresskv 2*512 | o 2*6144] f32 (~60KB/layer). The conv helper
reads the [C,2] history window (zeros at sequence start, kv_head==0), builds
xx = [hist ++ x], and writes the last two columns back each ubatch — the concat
naturally handles both prefill chaining and the T==1 shift. Read precedes write
in graph order; the fixed-offset copy is graph-reuse safe.
Verified: prefill anchors unchanged (bit-path identical, zero-history branch);
-ub 1 token-by-token run matches the golden reference at t4 (qlora_conv 0.084,
R_block 0.008 rel; attn_out 0.15 on one channel = f16 KV-cache rounding, washes
out by post-norm); final logits differ from full-batch only by a common-mode
shift that softmax cancels. Chat-template greedy smoke: think-block repetition
is gone — clean structured CoT and correct answer.
v0 limits documented in the helper: one state slot (single sequence); cache
rewinds leave the state stale.
* openpangu: NextN/MTP speculative decoding — 1.7-1.8x TG on CPU
Wire the three NextN layers (46-48) into ik's MTP speculative framework
(--spec-type mtp). v0 drafts with head 1 (layer 46), self-chained by the
framework.
- llama.cpp: add OPENPANGU to the cparams.mtp arch allowlist (it was silently
zeroed, which left the target context without a logits buffer once the server
enabled embeddings -> GGML_ASSERT(lctx.logits) in speculative_is_compat).
- load-tensors: MTP layers carry no mHC tensors (tail_use_mhc=false in the
reference) — create them only for base layers. nextn.* tensors were already
wired by the Stage-1 probe.
- build_openpangu: extract the attention sublayer into
build_openpangu_attention (shared base/MTP); add build_openpangu_mtp:
eh_proj(cat(enorm(embed), hnorm(prev_hidden))) -> one plain-residual Pangu
block (sandwich norms, convs+param_sink, MoE+shexp, no mHC/block_post_norm)
-> shared_head norm+head. MTP branch returns the draft graph when
mtp_op_type != NONE; main graph keeps all-token outputs under cparams.mtp.
MTP convs run batch-local (no conv-state slot) — affects acceptance only.
A/B (Q4_K_M, CPU, greedy, 192-token chat CoT completion, warm back-to-back,
medians of 3, bracketed B/A/B):
no-spec: 2.44 t/s (2.34-2.86)
--spec-type mtp:n_max=3: 4.23 / 4.49 t/s (brackets) => ~1.7-1.8x
Draft acceptance 34% on CoT prose (46% on repetitive text); spec and no-spec
greedy outputs are byte-identical. Headroom: conv-state for MTP drafts, n_max
tuning, true 3-head chaining (spec_step_idx).
* server: include draft_n/draft_n_accepted in /completion timings
get_formated_timings() (the /completion path) omitted the speculative
counters that get_timings() (the OAI path) already reports; add them,
guarded by n_draft_total > 0 like the OAI path.
* openpangu: position-indexed MoME conv-state ring — rollback-safe spec decoding + MTP draft chaining
The v0 single-slot conv state held the last-2 pre-conv latents of the most
recent batch, so any speculative draft rejection left latents of REJECTED
positions in the state and every later decode ran with wrong t-1/t-2 taps
(3 conv sites x 46 layers). At 192-token greedy runs every spec config
diverged from no-spec, each differently (rejection-pattern dependent).
Replace it with a per-layer ring cache_s_l [n_lora_q+n_lora_kv+n_head*v_dim, 16]:
column pos%16 holds position pos's pre-conv latents ([qa|ckv|o] packed).
Invariant: reads target only positions before the first batch token, which
are committed, and committed latents depend only on the committed prefix -
rollback-safe by construction, no checkpointing. Writes cover the last
min(T,16) batch positions in <=2 contiguous cpy segments; the copy sources
are views of the [hist ++ x] concat so the history read is an ancestor of
every write (read-before-write by graph dependency).
The ring is also allocated for the NextN/MTP layers, so the draft head
chains real conv taps across WARMUP -> sequential DRAFT_GEN steps (was
batch-local zero-history per draft token).
graph_reuse is forced off for the arch: ring view offsets are position-
baked and the reuse patcher only updates the standard K/V-store copies.
Measured cost on the CPU server path: none visible. ggml_set_rows driven
by an input index tensor is the future reuse-safe shape.
Verified (Q4_K_M, CPU, greedy 192-tok chat-CoT, warm single process):
- no-spec output byte-identical to pre-ring build
- spec output byte-identical to no-spec below the n_predict cap, for all
of n_max in {1,2,3,4,6} x p_min in {0,0.3,0.6} (old build: all diverged)
- acceptance n3-p0: 33.9% -> 60.9%; n3-p0.3: 58.1% -> 68.9%
- TG medians: no-spec 3.19-3.32 t/s; mtp:n_max=3,p_min=0.3 6.97 t/s (~2.1x)
* openpangu: DSA lightning indexer + SWA schedule — long-context correctness past the dense fallback
The dense fallback was exact only <=512 tokens (SWA window). This wires the real
DSA/SWA hybrid schedule, self-contained from GGUF keys the converter already
writes (openpangu.swa_layers + sliding_window_list; absent keys keep the old
dense fallback):
- SWA layers (30 base @512): the generic inp_KQ_mask_swa path, per-layer mask
choice in the builder. The NextN/MTP layers are SWA @2048 in the checkpoint
schedule; MTP graphs run in their own context, so the mask fill picks
hparams.n_swa_mtp when built with an MTP op type.
- DSA layers (16, every 3rd): lightning indexer implemented in-graph from the
Infer reference semantics (jointfix _pangu_torch_calib): q_idx = wq_b on the
post-conv post-norm q-lora latent (24x128), k_idx = rms-normed wk(x) shared
across heads, both NEOX-roped on the FIRST n_rot channels; score =
sum_g w_g * relu(q_g . k) in f32, causal-masked, exact top-k via
argsort + ggml_set_rows scatter into a -1e30 base -> additive selection mask
on the existing manual soft_max seam. Selection engages only when the causal
window exceeds index_top_k (2048); below that the layer is exactly dense.
- Indexer keys cached per position (cache_idx_l, f32 [128, kv_size], DSA layers
only) with the same committed-position invariant as the conv-state ring, so
speculative rollbacks stay safe.
- param sinks remain outside both the window and the selection budget, matching
the reference.
Verified (Q4_K_M, CPU):
- <=512 tokens: byte-identical to the dense build (96/160-token greedy)
- indexer scores vs a GGUF-dequant golden reference at 2101 tokens: 1e-3 rel
(f16 weight rounding); top-3 selection indices exact on all compared queries
- >512 coherence clean; 3.4K-token needle retrieval through active selection
(needle outside every SWA window, ~1300 positions pruned) answers exactly
* openpangu: MLA-latent KV cache — attention absorbed into the 512-latent, 14x smaller cache, ~2.2x TG
Store per position only [ckv_norm 512 | roped k_pe 64] (f32, k_l) plus the
transposed 512-latent (f32, v_l, v_trans layout); per-head K/V are never
materialized. q_nope is absorbed through attn_k_b (loaded 2D from the
converter split for base layers; derived at load via llm_prepare_mla for the
NextN layers - now guarded for layers without attention weights, e.g. the
idle NextN heads 2/3). The value side is the latent itself, up-projected
through attn_v_b after the weighted sum, matching the Infer _forward_dsa
reference. param sinks are native latent-space entries, which removes the
per-step full-cache concat+cast that dominated long-context decode.
llama_state row sizes now come from llama_kv_k_row_embd/llama_kv_v_row_embd
(arch-aware), fixing an out-of-bounds crash in the server prompt-cache save
path (hparams-derived 9216-wide rows vs actual 576-wide latent rows).
Verified (Q4_K_M, CPU): layer-0 attention output matches an f32 golden
reference computed from the same GGUF weight encodings (~1e-2 on O(1)
values); MTP spec output byte-identical to no-spec; 3.4K needle retrieval
through active DSA selection exact under greedy. Output differs from the
materialized build at the token level because attn_k_b/attn_v_b are
independently quantized tensors - both are legitimate Q4-fidelity encodings.
Perf (CPU, warm): no-spec TG 3.2-3.3 -> 6.9-7.1 t/s; mtp:n_max=3,p_min=0.3
-> 11.1 t/s (byte-exact, 67% acceptance); prefill 30.5 t/s at 3.4K; KV self
size at 4K ctx: 5.5 GiB -> 391 MiB. Not yet supported on the latent cache:
K-shift/defrag (context shifting) - unreached in current usage.
* openpangu: fence unsupported serving modes, truth-pass comments, drop dead weight/keys
Post-audit hardening. The cache's position-indexed side state (MoME conv ring,
DSA indexer keys) made several generic serving paths silently unsound; they are
now fenced loudly instead of documented as unsupported:
- s_l_position_ring flag on llama_kv_cache: the qnext-state predicate no longer
claims the conv ring, so per-seq state save, seq_cp and the s_copy graph skip it
- state save/restore refused for the arch at every llama_state_* entry (the ring
and idx_l are not in the state format; restoring without them diverges silently)
- K-shift/self-extend assert, defrag skips with a warning, server ctx_shift off
via new llama_model_supports_ctx_shift()
- single sequence enforced at context creation (n_seq_max > 1 refused)
- server prompt-cache reuse limited to pure extension via new
llama_model_supports_partial_kv_reuse(): mid-cache divergence reprocesses from
scratch (the 16-column ring cannot rewind); multi-turn continuation stays fast
- MTP draft length clamped to 13 via new llama_model_max_draft_tokens() so a
rejected draft can never overwrite the ring columns the next decode reads
- K/V cache types forced to f32 for the arch so the KV size log reports the truth
- cache_size(): real latent-cache branch (was falling through to the ~14x larger
materialized estimate used for offload planning)
- unused fused wkv_b no longer loaded (TENSOR_SKIP; the graph runs entirely on the
pre-split k_b/v_b), llm_prepare_mla openPangu special-case removed (it was a no-op)
- stale v0 comments rewritten to describe the shipped graph; converter stops
writing dead keys (dsa_layers, block_post_layernorm_idx) and the tokenizer
pre-hash is registered in convert_hf_to_gguf_update.py
Gates on this build: greedy spec output byte-identical to no-spec (EOS-terminated,
sha-equal); 3.4K needle retrieved exactly; -np 2 / state save / n_max=20 / stale
prefix reuse all refused or clamped with clear messages.
* openpangu: assert kv_head == first batch position at graph build
The ring, indexer and latent stores are addressed by absolute position through
kv_head; the fences make append-only decode the only reachable mode, but the
invariant was unchecked. Assert it at both graph entries (base and MTP) so any
future cache plumbing that breaks it fails at build instead of corrupting
output. Worst-case measurement builds pass pos = null and are exempt.
* openpangu: cont h_pre before the mHC broadcast mul (CUDA binbcast misreads strided views)
h_pre is a row-slice view of the fused mixes tensor. The CPU mul handles the
strides; the CUDA broadcast path reads the view as if contiguous, so token 0
mixes correctly and every later token gets h_post/h_res rows instead. First
divergent node in the whole graph (oracle rel 0.36 at opg0_attn_mhcpre_x,
fixed to 7.5e-5). Sibling views h_post/h_res were already cont-wrapped, which
is why only h_pre was exposed.
* openpangu: keep DSA zero-trick sources finite (CUDA clamp propagates the 0*(-inf) NaN)
The selection-mask base and zeros were built by scaling the MASKED scores by
zero, but post-mask sc contains -inf and 0 * -inf = NaN. The CPU clamp launders
NaN back to -1e30 (fminf/fmaxf ignore NaN); the CUDA clamp propagates it, so
every DSA layer emitted NaN masks at n_kv > top_k and logits collapsed
(observed: eval-callback CLAMP sum -1.3e36 on CPU vs nan on CUDA, 11748 NaNs
downstream). Scale the pre-mask finite scores instead, which is correct on any
backend regardless of clamp NaN semantics. Also defensively cont the strided
KQ_mask slice feeding the score add (same strided-view kernel class as the mHC
h_pre fix; unproven here but cheap). Gates after fix: 2600-token probe coherent,
3.4K needle exact ('7391') with and without MTP speculation, PP ~120 t/s.
* openpangu: f16 latent KV cache option (explicit -ctk/-ctv f16 halves cache memory, f32 stays default)
Track explicit cache-type requests through CLI/env; openPangu resolves no-request
to f32 (unchanged), accepts explicit f32/f16, warns and falls back to f32 for
BF16/quantized. Sink and cached-token KQ paths stay separate until after KQ so
the latent cache is read directly without the f32-only concat; value is the sum
of the sink and cache matmuls. Ring and DSA indexer caches stay f32; cache_size()
follows the resolved types.
* openpangu: enable graph reuse
* openpangu: wire multi-head MTP drafting
* openpangu: add MTP heads override
* openpangu: keep MTP update logits last
* openpangu: scope MTP warmup heads
* speculative: apply per-request MTP heads before warmup
* openpangu: fix multi-head MTP warmup computing on unwritten inputs
Each chained head called the build_inp_* helpers itself, so the warmup and
update graphs held one inp_tokens/inp_pos/inp_out_ids/KQ_mask tensor per
head while llama_set_inputs only fills the tensors the lctx pointers
reference, i.e. the last head's copies. Every head but the last read
unwritten compute-buffer memory: with heads=3 active even head 1's ring,
latent cache, and cached one-token draft were computed from garbage, which
is why depth-1 acceptance measured 4% against 98% for the heads=1 control.
Create the batch inputs once in build_openpangu and pass them to every
build_openpangu_mtp call, and fix the two chaining errors that were hiding
behind the garbage inputs:
- Shift the chained hidden: head k+1's row at position p consumes head k's
output row at p-1, the same convention head 1 uses for the target's
conditioned hidden rows. The predecessor of a batch's first row lives in
the previous warmup/update, carried across decodes through a new
inp_mtp_carry input backed by lctx.mtp_carry (written back per ubatch,
zeroed when a prompt warmup restarts from position 0).
- Fill head 3's cache row at draft step 2: each draft step runs one head,
so head 3's own decode at step 3 attended over a never-written row at
the step-2 position. Pre-write it from the committed carry.
Also include the active head count in the graph-reuse key next to the
existing step index (reuse stays forced off for this arch).
* speculative: default MTP drafting to a single head
A stage without an explicit heads= override previously resolved to 0,
meaning all model heads, so multi-head drafting was silently on by
default for models that carry more than one NextN layer. Keep it opt-in
(heads=N or heads=0 for all) until multi-head measures a win over the
single-head config; single-head models are unaffected either way.
* speculative: fence MTP head upshift over a warmed prefix
Deeper NextN heads only hold valid cache rows for spans that were warmed
with them. A request drafting with more MTP heads than the cached prefix
was warmed with (e.g. a heads=1 conversation continued with heads=3, a
pure extension the divergence fence deliberately allows) would read
never-written deeper-head rows: verification keeps the output correct,
but acceptance quietly collapses and any measurement taken there is
misleading.
Track the minimum head count the committed context has been warmed with
since position 0 and have the server reprocess from scratch when a
request asks for more. Also announce the model's NextN head count and
the single-head default once at MTP context setup.
* openpangu: skip dead MTP chain compute and stall-free carry readback
The update chain's last head and the draft-time row fill only matter for
their latent-cache and conv-ring writes; their FFN, norms, and shared
head fed nothing. Add a cache-writes-only mode to the MTP block builder
that returns after the attention block, and use it at both sites.
The multi-head carry readback previously synchronized the scheduler
after every warmup/update decode, a hard stall on CUDA. Issue the
device-to-host copy async on the backend stream instead (stream order
protects the source buffer from later graphs) and synchronize lazily
when the host buffer is next consumed or resized.
* openpangu: stop emitting fused kv_b tensor
* openpangu: default latent cache to f16
* openpangu: refuse unsupported latent cache types
* Window OpenPangu SWA cache reads
* Gather OpenPangu DSA decode reads
Gather DSA decode attention over the selected latent rows for OpenPangu base-model decode and verify graphs. The gathered branch now uses ggml_top_k order directly, runs maskless softmax over sinks plus selected rows for T <= 14, and derives values from the gathered k_l rows instead of the transposed latent cache.
* Chunk OpenPangu indexer prefill scoring
* Chunk OpenPangu prefill attention
* Gather OpenPangu sparse prefill attention
* Drop OpenPangu value cache
* Add OpenPangu indexer cache type flag
* Add OpenPangu q8_0 latent cache type
Store the OpenPangu MLA latent K cache as q8_0 via -ctk q8_0 (about 0.53x of
f16); the default stays f16 so behavior is unchanged without the flag. Latent V
stays f16/f32.
The q8 latent cache is a storage format only: it is dequanted to F32 before all
compute. K reads go through openpangu_build_k_latent_for_read, V derivation
through openpangu_build_v_latent_from_k (full 576-wide row to F32, then slice),
and the DSA gather paths already dequant via get_rows. Feeding a q8 latent view
directly into the KQ mul_mat corrupts large-context prefill, so that path is
removed for quantized caches. The cache write stages ckv and kpe through F32 and
writes one full 576-wide q8 row per token.
Verified on a small discriminator model: the default f16 path is byte-identical
to the prior code; the first-DSA-layer attention envelope is within 0.6% of the
f16 cache (linf_rel 0.0057); top-k selection is bit-identical between cache
types; the q8 latent cache is 0.531x the f16 size at 8K and 32K context; and
generation stays coherent on both the dense and DSA-gather paths at all tested
context lengths.
* Remove OpenPangu debug trace env knobs and redundant DSA_TOPK override
Drop the five LLAMA_OPENPANGU_*_TRACE debug-logging knobs (DSA_GATHER_TRACE,
IDX_CHUNK_TRACE, ATT_CHUNK_TRACE, PREFILL_GATHER_TRACE, SWA_WINDOW_TRACE) and the
LLAMA_OPENPANGU_DSA_TOPK override, which duplicated the -dsatk / --dsa-top-k CLI
flag; top-k now comes solely from cparams.dsa_top_k. The five perf-tuning knobs
(DSA_GATHER, IDX_CHUNK, ATT_CHUNK, ATT_KQ_MAX_MIB, PREFILL_GATHER) are retained
pending the perf battery. No change to default behavior.
* Subchunk OpenPangu DSA prefill gather to fit CUDA grid limit
The prefill gathered-attention ggml_get_rows produced dst rows = topk *
token_chunk (2048 * 256 = 524288) mapped to the CUDA grid.y dimension, which
caps at 65535, crashing with GET_ROWS invalid argument at long context (N_KV
around 10.5K with the natural topk of 2048). Split the prefill gather into token
subchunks so topk * subchunk_tokens stays within the grid limit, and guard the
decode gather with the same fit check (falling back to the dense masked path if
a pathological topk would not fit). The subchunking is over the token dimension
only, so per-token attention is unchanged and the result is numerically
identical. Verified: the GPU sweep runs past the old crash boundary to 22K+ with
zero CUDA errors; CPU and -ctk q8_0 paths unaffected.
* openpangu: fix scheduler node budget for chunked DSA prefill; drop unused attn_kv_b; remove env tunables
- Size the scheduler graph node budget for the chunked DSA prefill so 32K/ub2048 no
longer trips the hash-set reservation assert; derive the extra budget from the
builder's chunk/top-k/window structure with a fixed safety margin.
- Remove LLAMA_OPENPANGU_* environment tunables from both the node-budget estimator
and build_openpangu.cpp; use fixed constants in both so they stay in sync.
- Converter: emit only the split attn_k_b/attn_v_b projections and drop the unused
fused attn_kv_b tensor.
* openpangu: restore DeepSeek converter kv_b; drop trace env + dead code; fix dense-fallback node budget
- convert_hf_to_gguf.py: restore fused attn_kv_b in DeepseekV2Model (shared
parent); openPangu subclass keeps split-only k_b/v_b. Stops newly-converted
DeepSeek GGUFs from failing to load.
- src/llama.cpp: remove LLAMA_GRAPH_REUSE_TRACE getenv, hit/miss counters, and
the unconditional destructor log (no getenv or behavior change for any arch);
node-budget estimator now covers the dense-fallback (n_swa==0) attention-chunk
loop while skipping absent idx/top-k terms, preserving a strict overcount;
remove unreachable openPangu split-cache block.
- src/llama-context.h: drop now-dead graph_reuse_hits/misses members.
- include/llama.h: move type_k/type_v/idx_type_k *_explicit bools to struct end
to avoid a mid-struct ABI shift for out-of-tree consumers.
- src/graphs/build_openpangu.cpp: replace vestigial env-struct singletons with
the OPENPANGU_* constants; drop a redundant Sinkhorn permute round-trip
(one transpose; greedy output verified byte-identical).
Decode output unchanged (byte-identical greedy generation verified); shared-file
changes are openPangu-gated or restore the pre-PR baseline.
* openpangu: chat-parser support (reasoning split + thinking toggle)
Two openPangu-only fixes, both gated on the arch-unique token
<|pangu_text_start|> so no other model's parsing changes.
- chat-diff-analyzer: add a workarounds entry that force-sets TAG_BASED
reasoning with an empty start and a </think> end. openPangu prefills
<think> in the generation prompt, so the output is delimited only by
</think>; the differential detector otherwise learns start="<think>"
from the assistant-history form and fails to split, leaking reasoning
into content. Same shape as the existing Laguna prefill patch.
- chat.cpp: bridge enable_thinking to the template's `thinking` variable.
openPangu's template gates reasoning on `thinking` rather than the
ecosystem-standard `enable_thinking`, so the standard toggle was inert.
An explicit `thinking` chat_template_kwarg still overrides via the
extra_context merge.
Blast radius: test-chat-auto-parser 437/437 unchanged; the sole
test-chat-template diff is a pre-existing GLM trailing-newline.
* openpangu: use ggml_cast for latent dequant reads
Replace ggml_cpy(view, ggml_new_tensor_2d(F32, ...)) with ggml_cast in the MLA
latent V-from-K and K-read helpers. ggml_cast emits the identical GGML_OP_CPY
node into a fresh f32 tensor, so behavior is unchanged; it is the idiomatic
form. Per review.
* openpangu: narrow SWA reuse-key fields to 32-bit
The openpangu_swa_window_view reuse key stored n_kv/n_tokens/window/pad as
int64_t, but these are bounded well under 2^31 (window/pad are uint32_t at
source; n_kv/n_tokens <= context length). Narrow to int32_t/uint32_t and drop
the widening casts. w_view/win_off stay int64_t: they feed ggml view
dims/offsets. Per review.
* openpangu: precompute param_sink derived tensors at load
The per-layer attention-sink block (sink_blk [576,NS]) and its transposed
latent (s_lat_t [NS,512]) are pure functions of the layer weights, yet were
rebuilt every eval across all 49 layers (RMS-norm + cast + concat + transpose).
Compute them once at load, mirroring the wk_b derived-weight precompute, and
read the stored tensors in build_openpangu_attention. Numerically identical;
removes per-token work at decode.
* openpangu: replace conv position-ring with ggml_ssm_conv + spec-rollback checkpoint
Migrate the MoME depthwise causal conv (three sites per attention sublayer:
qa-lora, compressed-kv, attn-out) from the bespoke 16-column position-indexed
ring onto the core ggml_ssm_conv op with a recurrent conv-state slot.
Cache: s_l becomes [2*conv_col_ne, qnext_state_slots], holding the (d_conv-1)=2
history taps per channel for the three sites (float offsets 0 / 2*n_lora_q /
2*(n_lora_q+n_lora_kv)). Drops the conv_hist_idx / conv_write_idx graph inputs
and their fill in llama_set_inputs; adds one single-sequence sq input for
ggml_ssm_conv shared across the three sites and the MTP head.
Speculative rollback: the position ring self-healed rejected draft columns by
absolute position; a recurrent slot does not, since seq_rm is a no-op for
recurrent state. openPangu is admitted at the three spec-checkpoint save/init
gates so the whole-slot shadow checkpoint (gpu-fallback) snapshots the conv
slot before drafting and restores it before the accepted-token replay. The
restore path is already keyed on ckpt.valid, so no gate change is needed there.
Per-step checkpoint mode is declined for openPangu, which has no SSM recurrent
term, so auto mode resolves to the whole-slot shadow.
Gated: non-spec needle unchanged; MTP-spec needle correct with healthy draft
acceptance (rollback verified via the acceptance canary).
* openpangu: single ggml_concat copy for the latent cache store
The non-quantized latent store split the [ckv | roped k_pe] row into two views
and two cache copies, with a base_offset field on the CacheCopy struct to place
the second one. Match the quantized path: concat the two parts and do one copy
into the cache row. This drops the second cache-copy slot (OPENPANGU_COPY_K_KPE)
and removes base_offset from CacheCopy entirely.
Cache contents are unchanged: the concat writes the same [ckv 512 | k_pe 64]
bytes to the same row. Gated on the needle for both the f16 latent path (the one
that changed) and the q8 latent path, plus coherence.
* openpangu: reuse the shared kr_l indexer cache instead of a separate idx_l
The DSA lightning indexer stored its per-position keys in an openPangu-only idx_l
cache, parallel to the kr_l indexer cache GLM-DSA already uses. Both have the same
storage contract: [indexer_head_size, kv_size], idx_type_k dtype, one row per KV
cell, written at kv_head and read [dim, n_kv] from zero. openPangu now allocates
its indexer keys into kr_l and shares the dsa_cache_copies graph-reuse fixup.
The fixup patch is factored into a helper that both the generic path and the
openPangu update_cache_copies branch call, so the openPangu indexer copy is
repointed to the current kv_head on graph reuse like every other cache write.
This drops the idx_l vector, its allocation and memory accounting, and the
openPangu third cache-copy slot (now one latent copy per layer).
Per-arch allocation predicates stay separate (GLM uses indexer_is_full, openPangu
uses the window==0 DSA schedule); only the kr_l storage and the copy fixup are
shared. openPangu keeps its no-shift/no-defrag/no-state-I/O behavior, and the GLM
Hadamard/k-shift logic stays GLM-gated.
Gated: needle correct on f16 and q8 latent caches and under MTP speculation
(acceptance unchanged at 0.67), plus coherence.
* openpangu: discard pos-0 graphs from reuse; retire stale conv-state comments
The ggml_ssm_conv refactor bakes the pos-0 conv-state reset into graph
topology (a scale-by-zero node on the state view). A graph built at pos 0
could be reused at pos > 0 when the batch shape and padded n_kv match (a
1-token prompt followed by TG is the concrete case), zeroing the conv
history on every reused decode. Admit openPangu at the existing
reset_previous gate so pos-0 graphs are discarded from reuse, the same
guard the qnext recurrent state relies on.
Also retire the internal phase-plan comments the conv refactor left
behind: they claimed the spec-checkpoint wiring had not landed in the
commit that landed it, and misdescribed the s_l slot as awaiting rollback
support.
Gated: needle 8457 on f16 and q8 latent, MTP-spec needle (drafts fully
accepted), coherence.
* openpangu: drop the _explicit cache-type plumbing; validate unconditionally
Review follow-up (item 1 of the second review). The explicit/default
distinction carried less than claimed: the latent K/V fallback was f16,
which is already the -ctk/-ctv and API default, so distinguishing unset
from set-to-the-default bought nothing, and the two bools were behaviorally
redundant. The only load-bearing use was the indexer cache, where openPangu
defaulted to f32 while -ictk defaults to f16. Gating the f16 indexer
directly (needle on f16 and q8 latent paths, MTP speculation, coherence)
shows no quality difference, so openPangu now takes the standard f16
indexer default and the f32 special case is gone. Default indexer cache
memory halves (64 -> 32 MiB at c 8192).
Removes type_k_explicit/type_v_explicit/idx_type_k_explicit from llama.h,
the cparams/mparams plumbing, and common; the resolve helpers become plain
unconditional validators, so -ctk q8_0 is honored and an unsupported type
errors out at load instead of silently coercing.
Gated: needle 8457 on the new f16-indexer default, on q8 latent with MTP
speculation, and with -ictk f32 explicitly honored (64 MiB f32 buffer in
the load log); -ctk q4_0 and -ictk q4_1 refused with a clear error.
* openpangu: keep MTP draft decodes position-contiguous under speculation
The MTP framework's one-token draft shortcut caches a prediction one row
past the accepted prefix during the accepted-token update, then skips
re-decoding the last sampled token at the next draft round. A
mask-addressed cache tolerates the resulting position gap; openPangu's
position-addressed append-only cache (cell == position) does not: after a
rollback the next draft decode lands one cell behind its position, and
after a full acceptance the cache head sits one row ahead of the next
draft base, either way tripping the kv_head == pos[0] invariant and
aborting the server. The checkpoint admission in the conv refactor made
this the standard openPangu speculative flow; the needle-first gates
never generated enough draft rounds against a short prompt to reach it.
Decline the shortcut re-seed for openPangu in mtp_accept_batch (restoring
the drafting behavior all measured acceptance numbers were taken on) and
trim rows at or beyond the draft base in mtp_speculative_gen_draft, so
every draft decode stays position-contiguous with the cache head.
Gated: the crashing flow (short prompt, 512-token spec generation, then a
second request) completes with acceptance 0.60 prose / 0.87 code,
matching the pre-checkpoint baseline profile; needle 8457 plus coherence
on f16+spec and q8+spec.
* openpangu: remove stale ring limits and fix MTP graph reuse
* cli: preserve speculative carry on fallback
Decode an already-emitted pending token when a draft cannot be used instead of sampling unchanged logits and duplicating output. Document single-head MTP as the default and multi-head modes as experimental.
---------
Co-authored-by: Joel Farthing <262452229+joelfarthing@users.noreply.github.com>
* Support MiMo DFlash draft conversion
* Fix MiMo2 DFlash capture row pruning
* Fix MiMo DFlash draft RoPE and value scale
* Honor partial_rotary_factor in DFlash draft RoPE dim count
The draft set rope.dimension_count to the full head_dim (128), ignoring the
MiMo DFlash draft's partial_rotary_factor=0.5. The correct count is
head_dim*partial_rotary_factor=64; the remaining dims are NoPE. With the full
head_dim the upper half of each head receives position rotation it was never
trained for, which roughly halves draft acceptance on code (~26% -> ~60% once
corrected). RoPE base (5e6) and value scale (0.612) were already correct.
* Filter weight-map shard discovery to files that exist
get_model_part_names_from_weight_map() returned shard names straight from the
index weight_map without checking they exist. A model dir with a stale
model.safetensors.index.json but no safetensors shards would then set
is_safetensors=True and skip the pytorch_model*.bin fallback, failing later when
opening the missing files. Filter to shards present on disk so a stale index
falls through to the other weight formats.
* DFlash: store backbone_rotary_base in dedicated GGUF key
backbone_rotary_base (the target model's RoPE theta used when encoding
context K/V) was written to rope.freq_base, clobbering the draft
model's own rope_theta. For MiMo this swapped 10000 → 5000000 in the
draft attention path.
Fix: write backbone_rotary_base to a dedicated dflash.backbone_rotary_base
GGUF key and read it into hparams.dflash_backbone_rotary_base. In
build_dflash_kv_cache, use target_freq_base (the new hparam when set,
falling back to freq_base) for the context-K RoPE call. The draft model's
own rope.freq_base is now set correctly from rope_theta.
Existing MiMo DFlash GGUFs must be reconverted.
---------
Co-authored-by: Joel Farthing <262452229+joelfarthing@users.noreply.github.com>
* Add llama.cpp changes for dots1 support
* Add python changes for dots1 support
* Fix to make it convert
* Remove V reshaping, remove BOS by default for dots1 and fix warmup to handle models without BOS
* Minor fix
* Remove commented lines
* lora : fix llama conversion script with ROPE_FREQS
* convert : refactor rope_freqs generation
This should also fix vocab-only conversion for Phi-3.
* convert : adapt MiniCPM3 to separate rope_freqs insertion
MiniCPM3's tokenizer is treated as a SentencePiece tokenizer to avoid
having to run its custom Python code which mixes tokenization
in the same file as tool calls.
gguf-py : add long and short RoPE factors to tensor mappings
Empty, but the key names are used to populate the mappings.
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: Francis Couture-Harpin <git@compilade.net>
* conflict resolution
* Changes to make work and add longrope support
* Changes to n_attention_wv rule
* Untested support of 253B
* DeciLMCausalModel now reads rope_theta from config.json properly
* Remove errant Granite mentions
* Better n_attention_vw rule
* Update vocab.py
---------
Co-authored-by: Yee Man Chan <ymchan@gmail.com>
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Deepseek MLA Optimizations
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
* Make MLA optional
* Remove some unnecessary copies in the MLA attention
* Deepseek MLA Optimizations V2 (#195)
* Avoid allocating MHA KV cache when MLA is turned on
* Added missing gguf-py file
* Added final optimizations
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
* Make sure we do have wk_b and wv_b before enabling MLA
---------
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Use type_k and type_v to set the types of the MLA caches
They were hard-coded at f16.
On my Ryzen-7950X with native bf16 support I get a fairly
significant PP performance boost with bf16 KV-cache:
PP-4096 = 320 t/s up from 292 t/s with fp16 KV-cache.
* Better gemm strategy when nth > nhead
It gives a ~10% PP performance boost for DeepSeek-Lite with 32 threads
(with or without MLA).
Before this commit, when nth > nhead heads were processed
sequentially with all nth threads participating in each
matrix multiplication. Now we ind the gcd of nhead and
nth and split threads into nth/gcd groups, each group
processing nhead/gcd heads.
---------
Co-authored-by: Saood Karim <saood05@gmail.com>
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Merging mainline - WIP
* Merging mainline - WIP
AVX2 and CUDA appear to work.
CUDA performance seems slightly (~1-2%) lower as it is so often
the case with llama.cpp/ggml after some "improvements" have been made.
* Merging mainline - fix Metal
* Remove check
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* feat: add changes to handle jina v2 base code
* fix: do not complicate things
* fix: fix the usage of the code model
* fix: fix comments
* fix: fix linting issues
* fix: remove ollama patches
* style : minor
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* common : increase max number of experts to 160
* common : add tensors ATTN_Q_A, ATTN_Q_A_NORM, ATTN_Q_B, ATTN_KV_A_MQA, ATTN_KV_A_NORM, ATTN_KV_B needed by DeepSeek-V2 MLA (multi-head latent attention) architecture
* common : add model header parameters: leading_dense_block_count, expert_feed_forward_length, expert_shared_count, expert_weights_scale, attention.q_lora_rank, attention.kv_lora_rank, rope.scaling.yarn_log_multiplier
* convert-hf : add model conversion support for DeepseekV2ForCausalLM
* llama : add model types for DeepSeek-V2 and DeepSeek-V2-Lite models
* llama : add two new llm_build_moe_ffn() arguments: scale_w (whether to scale weights of selected MoE experts) and w_scale (numerical value of the scaling factor)
* llama : add inference support for LLM_ARCH_DEEPSEEK2
---------
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
* common : increase max number of experts to 128
* common : add tensor LLM_TENSOR_FFN_NORM_EXPS for normalization before MoE that runs in parallel to attention + ffn
* gguf-py : add architecture-specific block mappings that override selected general block mappings
* convert-hf : add model conversion support for ArcticForCausalLM
* convert-hf : use added_tokens_decoder from tokenizer_config.json to redefine tokens from SentencePiece model (only for ArcticForCausalLM)
* llama : add inference support for LLM_ARCH_ARCTIC
---------
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
* feat: first things to do
* feat: create tensors for Jina architecture
* fix: use other tensors
* feat: embedding gets results
* fix: fix usage of ALIBI
* fix: clean prints
* fix: do some cleanup unused vars
* fix: revert changes to Makefile and CMakeLists
* fix: revert some changes
* fix: fix small detail
* fix: fix convert formatting
* fix: fix linting and editor
* feat: set proper vocab settings
* fix: JinaBertForMaskedLM registration
* feat: support q_normalization and k_normalization in Jina arch
* feat: handle gpt2 tokenizer with Jina architecture
* feat: example comments in embedding
* feat: rename Jina Bert to Jina Bert V2
* fix: add some changes as per review
* feat: proper KQ_pos for Jina embeddings
* feat: add capacity to load models ES and DE for Spanish
* llama : fix pre-tokenizers
* ggml : full ALiBi support
* ggml : update ggml_soft_max_ext() CUDA, SYCL
* ggml : ggml_flash_attn_ext() support ALiBi (CPU)
* ggml : ggml_flash_attn_ext() support ALiBi (Metal)
* ggml : fix warning
* ggml : ggml_flash_attn_ext() support ALiBi (CUDA)
ggml-ci
* minor : clean-up
* embedding : add warning about missing SEP
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* support qwen2moe
* fix-review
* metal : support unary ops for nelements % 4 != 0
* metal : require contiguousness for float4 unary kernels
* metal : require contiguousness for float4 unary kernels (cont)
* fix-review
* names : for brevity "SHARED_EXP" -> "SHEXP"
* llama : reuse build_moe_ffn()
* llama : add model type name
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Add Command R Plus GGUF
* Add Command R Plus GGUF
* Loading works up to LayerNorm2D
* Export new tensors in 1D so they are not quantized.
* Fix embedding layer based on Noeda's example
* Whitespace
* Add line
* Fix unexpected tokens on MPS. Re-add F16 fix. ((Noeda)
* dranger003: Fix block index overflow in CUDA dequantizing.
* Reverted blocked multiplication code as it still has issues and could affect other Llama arches
* export norms as f32
* fix overflow issues during quant and other cleanup
* Type convention
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* dranger003: Fix more int overflow during quant.
---------
Co-authored-by: S <seast@Ss-Mac-Studio.local>
Co-authored-by: S <s@example.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* initial commit for sealion support
* add sealion support
* minor fix
* q/k ln and pos_embd only if required
* Apply suggestions from code review
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* minor : clear whitespaces
---------
Co-authored-by: bryan <bryansiow@aisingapore.org>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* ggml : update mul_mat_id to use the same tensor for all the experts
* update cuda
* minor
* update metal
* update test-backend-ops
* fix cuda
* Update ggml-metal.m
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* update convert.py
* update convert-hf-to-gguf.py
* update convert.py for mixtral hf models
* Update convert-hf-to-gguf.py
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* cuda : support non-pow-2 number of experts
* allow quantize to work for split and merged experts models in the same way
* cleanup + disable mmap automatically with split tensors models
* update imatrix
* test-backend-ops : test qwen argsort
* update grok model loading
* llama : add merged experts tensors to the grok tensor map
* minor
* gguf : bump version
* fix quantizing of merged experts
* convert-hf-to-gguf.py : update grok (untested)
* make linter happy
* cuda/argsort : use shared memory instead of pool memory
* convert : fix grok tensor names
* metal : add support for non-pow-2 argsort
* llama : more loader cleanup, better error checking
* cuda : fix warning
* llama : still use mmap for loading old models, but copy the data to a host buffer
* add review note
* llama : remove ffn tensor counting + add sanity check
ggml-ci
* convert : fix handling of n_experts == None
ggml-ci
* imatrix : fix ncall counters
* llama : produce error if imatrix size does not match
* quantize : terminate on errors + trace logs
ggml-ci
* metal : pad shared memory to 16 bytes
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* mamba : begin working on support for Mamba SSM
* mamba : begin figuring out how to (ab)use the kv cache for Mamba
* mamba : recurrent inference almost works, but incoherent
* mamba : recurrent inference WORKS!!!
* convert : optionally use d_conv and d_state from config.json for Mamba
* mamba : refactor recurrent conv, resulting in 20% perf increase
It's still slower than I'd like, but I did not really optimize `ggml_exp` yet.
I also refactored `ggml_exp` to work with tensors with more than 2 dimensions.
* ggml : parallelize ggml_exp
This results in 8% faster token generation for Mamba-130M.
* mamba : simplify the conv step with a self-overlapping view
Turns out the conv_state can be made smaller by one column.
Note that this breaks existing GGUFs of Mamba,
because the key_value_length field is tied to the conv_state size.
Convolution with a self-overlapping view is cool!
And it's much simpler than what I initially thought would be necessary
to make the convolution step work with more than 1 token at a time.
Next step is to make the SSM step work on batches of tokens too,
and thus I need to figure out a way to make a parallel selective scan
which will keep the ssm_state small and won't make it bigger
by a factor of (n_layer * batch_size).
* llama : fix Mamba KV self size wrongly displaying as f16 instead of f32
Relatedly, I also tried to see if other types than f32 worked for the states,
but they don't, because of the operators used.
It's probably better anyway to keep lots of precision there,
since the states are small anyway.
* mamba : fix self-overlapping view depth stride
* mamba : handle batches of more than 1 token
This means running Mamba no longer crashes when using the default settings!
And probably also slightly faster prompt processing.
Both batched and non-batched processing yield the same output.
Previously, the state was not cleared when starting a sequence.
Next step is to make the KV cache API work as expected for Mamba models.
* ggml: add ggml_ssm_scan to help with parallel selective scan
If the selective scan was implemented without a custom operator,
there would be waaay too many nodes in the graph. For example,
for Mamba-130M, with a batch size of 512 (the default),
a naive selective scan could add at least 24*512=12288 nodes,
which is more than LLAMA_MAX_NODES (8192),
and that's only for the smallest Mamba model.
So it's much cleaner with a custom operator.
Not sure about the name, though.
* ggml : in ggml_ssm_scan, merge multiple rows in the same vec operation
This will help with performance on CPU if ggml_vec_mul_f32
and ggml_vec_add_f32 are ever optimized with SIMD.
* mamba : very basic quantization support
Mostly works, but there is currently no difference
between the variants of a k-quant (e.g. Q4_K_S and Q4_K_M are the same).
Most of the SSM-specific weights can be kept in f32 without affecting
the size that much, since they are relatively small.
(the linear projection weights are responsible for most of Mamba's size)
Too much quantization seems to make the state degrade quite fast, and
the model begins to output gibberish.
It seems to affect bigger models to a lesser extent than small models,
but I'm not sure by how much.
Experimentation will be needed to figure out which weights are more important
for the _M (and _L?) variants of k-quants for Mamba.
* convert : fix wrong name for layer norm weight of offical Mamba models
I was using Q-bert/Mamba-* models before, which have a slighlty different
naming scheme for the weights.
(they start with "model.layers" instead of "backbone.layers")
* mamba : fuse more steps of the SSM scan in the ggml_ssm_scan operator
This increases performance on CPU by around 30% for prompt processing,
and by around 20% for text generation.
However, it also makes the ggml_exp and ggml_soft_plus operators unused.
Whether or not they should be kept will be decided later.
* convert : for Mamba, also consider the "MambaLMHeadModel" arch name
It's the name of the class of the official implementation,
though they don't use it (yet) in the "architectures" field of config.json
* mamba : fix vocab size problems with official models
The perplexity was waaaay to high for models with a non-round vocab size.
Not sure why, but it needed to be fixed in the metadata.
Note that this breaks existing GGUF-converted Mamba models,
but **only if** the vocab size was not already rounded.
* ggml : remove ggml_exp and ggml_soft_plus
They did not exist anyway outside of this branch,
and since ggml_ssm_scan fused operations together, they are unused.
It's always possible to bring them back if needed.
* mamba : remove some useless comments
No code change.
* convert : fix flake8 linter errors
* mamba : apply suggestions from code review
* mamba : remove unecessary branch for row-wise ssm_state and C multiplication
It was previously done to avoid permuting when only one token is processed
at a time (like when generating text), but permuting is cheap,
and dynamically changing the compute graph is not future-proof.
* ggml : in ggml_ssm_scan, use more appropriate asserts
* ggml : rename the destination pointer in ggml_compute_forward_ssm_scan_f32
* mamba : multiple sequences, but one at a time
This is a step towards making this Mamba implementation usable
with the server example (the way the system prompt is kept when clearing
the client slots will need to be changed before this can work, though).
The KV cache size for this kind of model is tied to the maximum number
of sequences kept at any single time.
For now, this number is obtained from n_parallel (plus one,
to have an extra sequence to dedicate to the system prompt),
but there might be a better way to do this which won't also
make the main example use 2 cells even if only 1 is really used.
(for this specific case, --parallel 0 helps)
Simultaneous sequence processing will probably require changes to
ggml_ssm_scan, and possibly a new operator for the conv step.
* mamba : support llama_kv_cache_seq_cp
This (mis)uses the logic around K shifts, because tokens in a state
can't be shifted anyway, and because inp_K_shift has the right shape and type.
Using ggml_get_rows is a nice way to do copies, but copy chains can't work.
Fortunately, copy chains don't really seem to be used in the examples.
Each KV cell is dedicated to the sequence ID corresponding to its own index.
* mamba : use a state mask
It's cleaner than the previous heuristic of
checking for the pos of the first token in the batch.
inp_KQ_mask could not be re-used for this, because it has the wrong shape
and because it seems more suited to the next step of
simultaneous sequence processing (helping with the problem of
remembering which token belongs to which sequence(s)/state(s)).
* llama : replace the usage of n_ctx with kv_self.size in many places
* mamba : use n_tokens directly instead of n_tok
* mamba : in comments, properly refer to KV cells instead of slots
* mamba : reduce memory usage of ggml_ssm_scan
From 290.37 MiB to 140.68 MiB of CPU compute buffer size
with Mamba 3B with a batch size of 512.
The result tensor of ggml_ssm_scan was previously a big part
of the CPU compute buffer size. To make it smaller,
it does not contain the intermediate ssm states anymore.
Both y and the last ssm state are combined in the result tensor,
because it seems only a single tensor can be returned by an operator
with the way the graph is built.
* mamba : simultaneous sequence processing
A batch can now contain tokens from multiple sequences.
This is necessary for at least the parallel example, the server example,
and the HellaSwag test in the perplexity example.
However, for this to be useful, uses of llama_kv_cache_seq_rm/cp
will need to be changed to work on whole sequences.
* ggml : add ggml_ssm_conv as a new operator for the conv step of Mamba
This operator makes it possible to use and update the correct states
for each token of the batch in the same way as ggml_ssm_scan.
Other solutions which use existing operators would need loops which would
add too many nodes to the graph (at least the ones I thought of).
Using this operator further reduces the size of the CPU compute buffer
from 140.68 MiB to 103.20 MiB with Mamba 3B with a batch size of 512.
And (at least on CPU), it's a bit faster than before.
Note that "ggml_ssm_conv" is probably not the most appropriate name,
and it could be changed if a better one is found.
* llama : add inp_s_seq as a new input tensor
The most convenient implementation to select the correct state (for Mamba)
for each token is to directly get the correct index from a tensor.
This is why inp_s_seq is storing int32_t and not floats.
The other, less convenient way to select the correct state would be
to have inp_KQ_mask contain 1.0f for each state used by a token
and 0.0f otherwise. This complicates quickly fetching the first used
state of a token, and is also less efficient because a whole row
of the mask would always need to be read for each token.
Using indexes makes it easy to stop searching when there are
no more sequences for a token, and the first sequence assigned
is always very quickly available (it's the first element of each row).
* mamba : support llama_kv_cache_seq_cp copy chains
* mamba : support shifting and dividing the kv cache pos
* mamba : make the server and parallel examples work with whole sequences
A seq_id is dedicated to the system prompt in both cases.
* llama : make llama_kv_cache_seq_rm return whether it succeeded or not
* mamba : dedicate an input tensor for state copy indices
This is cleaner and makes it easier to adapt when/if token positions
(and by extension, inp_K_shift) are no longer integers.
* mamba : adapt perplexity, batched, and batched-bench examples
* perplexity : limit the max number of sequences
This adapts to what the loaded model can provide.
* llama : add llama_n_max_seq to get the upper limit for seq_ids
Used by the perplexity example.
* batched : pass n_parallel to the model's context params
This should have been there already, but it wasn't.
* batched-bench : reserve sequences to support Mamba
* batched-bench : fix tokens being put in wrong sequences
Generation quality isn't what's measured in there anyway,
but at least using the correct sequences avoids using non-consecutive
token positions.
* mamba : stop abusing attention metadata
This breaks existing converted-to-GGUF Mamba models,
but will allow supporting mixed architectures like MambaFormer
without needing to break Mamba models.
This will also allow changing the size of Mamba's states
without having to reconvert models in the future.
(e.g. using something else than d_conv - 1 columns for the conv_states
will not require breaking existing converted Mamba models again)
* gguf-py : add new KV metadata key-value pairs for Mamba
* llama : add new metadata key-value pairs for Mamba
* llama : guard against divisions by zero when n_head is 0
* mamba : rename "unlimited" KV cache property to "recurrent"
* mamba : more correctly update the "used" field of the KV cache
* ggml : in ggml_ssm_scan, use a threshold for soft_plus
This is how the official Mamba implementation does it,
and it's also what torch.nn.Softplus does.
* convert : for Mamba, fallback to internal NeoX tokenizer
The resulting models are exactly the same
as if the tokenizer.json and tokenizer_config.json of GPT-NeoX were there.
* mamba : support state saving and restoring
* ggml : implicitly pass src tensors through dst for Mamba-related ops
* mamba : clarify some comments
* server : fix cache_tokens not getting correctly resized
Otherwise, when the "we have to evaluate at least 1 token" special case
was triggered, an extra token was kept in cache_tokens even if it was
removed from the KV cache.
For Mamba, this caused useless prompt reprocessing when the previous
request triggered the above case.
* convert-hf : support new metadata keys for Mamba
For the models available at
https://huggingface.co/collections/state-spaces/transformers-compatible-mamba-65e7b40ab87e5297e45ae406
* mamba : rename metadata to be more similar to transformers library
This breaks existing converted-to-GGUF models,
but the metadata names are more "standard".
* mamba : support mamba-*-hf models
These models share their token_embd.weight with their output.weight
* mamba : add missing spaces
This is purely a formatting change.
* convert-hf : omit output.weight when identical with token_embd.weight
Only for Mamba for now, but it might be relevant for other models eventually.
Most Mamba models actually share these two tensors, albeit implicitly.
* readme : add Mamba to supported models, and add recent API changes
* mamba : move state_seq and state_mask views outside layer loop
A few tensors were also missing `struct` in front of `ggml_tensor`.
* BERT model graph construction (build_bert)
* WordPiece tokenizer (llm_tokenize_wpm)
* Add flag for non-causal attention models
* Allow for models that only output embeddings
* Support conversion of BERT models to GGUF
* Based on prior work by @xyzhang626 and @skeskinen
---------
Co-authored-by: Jared Van Bortel <jared@nomic.ai>
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* convert : support Mixtral as LLAMA arch
* convert : fix n_ff typo
* llama : model loading
* ggml : sync latest ggml_mul_mat_id
* llama : update graph to support MoE
* llama : fix cur -> cur_expert
* llama : first working version
* llama : fix expert weighting in the FFN
* ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only)
* ggml : add n_as argument to ggml_mul_mat_id
* ggml : fix ggml_get_rows to take into account ne02 / ne11
* metal : add more general support for ggml_get_rows + tests
* llama : add basic support for offloading moe with CUDA
* metal : add/mul/div use general kernel when src1 not cont
* metal : reduce the kernel launches for ggml_mul_mat_id
* ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D
* ggml : update get_rows f16 and q
* cuda : support non-contiguous src1 in get_rows
* llama : offload missing ffn_moe_silu
* metal : fix ggml_get_rows to work with non-cont src1
* metal : add indirect mat-vec kernels for all quantization types
* llama : do not quantize expert gating tensors
* llama : add n_expert and n_expert_used to hparams + change quants
* test-backend-ops : add moe test
* cuda : fix get_rows when ncols is odd
* convert : determine n_ctx correctly
* metal : fix ggml_mul_mat_id for F32
* test-backend-ops : make experts more evenly probable (test_moe)
* test-backend-ops : cleanup, add moe test for batches
* test-backend-ops : add cpy from f32 -> all types test
* test-backend-ops : fix dequantize block offset
* llama : fix hard-coded number of experts
* test-backend-ops : simplify and disable slow tests to avoid CI timeout
* test-backend-ops : disable MOE test with thread sanitizer
* cuda : fix mul_mat_id with multi gpu
* convert : use 1e6 rope_freq_base for mixtral
* convert : fix style
* convert : support safetensors format
* gguf-py : bump version
* metal : add cpy f16 -> f32 kernel
* metal : fix binary ops for ne10 % 4 != 0
* test-backend-ops : add one more sum_rows test
* ggml : do not use BLAS with ggml_mul_mat_id
* convert-hf : support for mixtral-instruct (#4428)
* convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct
* convert : use sentencepiece tokenizer for Mixtral-instruct
* convert : make flake8 happy
* metal : fix soft_max kernels
ref: 1914017863
* metal : limit kernels to not use more than the allowed threads
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Radek Pilar <github@mrkva.eu>
* gguf-py: Refactor and add file reading support
* Replay changes from #3871
Credit to @cebtenzzre for that pull
* Various type annotation fixes.
* sort imports with isort (again)
* Fix missing return statement in add_tensor
* style cleanup with flake8
* fix NamedTuple and Enum usage
* Fix an issue with state init in GGUFReader
Move examples to an examples/ directory
Clean up examples
Add an example of modifying keys in a GGUF file
Update documentation with info on examples
Try to support people importing gguf/gguf.py directly
* Damagage is not a word.
* Clean up gguf-py/examples/modify_gguf.py whitespace
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
* Update gguf-py/examples/modify_gguf.py formatting
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
* Update gguf-py/gguf/gguf_reader.py type hint
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
* Make examples executable, formatting changes
* Add more information to GGUFReader and examples comments
* Include a gguf Python package version bump
* Add convert-gguf-endian.py script
* cleanup
* gguf-py : bump minor version
* Reorganize scripts
* Make GGUFReader endian detection less arbitrary
* Add JSON dumping support to gguf-dump.py
Which I kind of regret now
* A few for gguf-dump.py cleanups
* Murder accidental tuple in gguf-py/scripts/gguf-dump.py
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
* cleanup
* constants : remove unneeded type annotations
* fix python 3.8 compat
* Set up gguf- scripts in pyproject.toml
* And include scripts/__init__.py, derp
* convert.py: We can't currently support Q8_0 on big endian.
* gguf-py: SpecialVocab: Always try available sources for special token ids
gguf-py: SpecialVocab: Try to load merges from merges.txt if not in tokenizer.json
gguf-py: SpecialVocab: Add 'add_bos_token' type bools to GGUF metadata
u
* cleanup
* Promote add_X_token to GGUF metadata for BOS and EOS
---------
Co-authored-by: Jared Van Bortel <jared@nomic.ai>
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
Kawrakow
Joel Farthing
Samuel Oliveira Alves
firecoperana
Thireus ☠
saood06
Kawrakow
Joan Fontanals
fairydreaming
Joan Fontanals
Georgi Gerganov
liuwei-git
Shijie
Pierrick Hymbert
Carolinabanana
bryanSwk
slaren
Julius Arkenberg
compilade
Sourab Mangrulkar
Jared Van Bortel
Douglas Hanley
Guoteng
chiranko
manikbhandari
Nam D. Tran
Shintarou Okada
Ebey Abraham
Kerfuffle