feat: import qwen4exp (Qwen3.8-Flash-Next) support from upstream PR #27742 - #324
Conversation
Adds LLM_ARCH_QWEN4EXP with its hparams and tensor loading. The graph comes in the next commit; this makes the model load and report correct metadata. - hyper-connections set n_embd_out_impl = hc_count * n_embd, so the residual stream is 4x wide and there is no output_norm: the final mixer's hc_norm is the last norm in the model. - registered as hybrid and given the same recurrent/attention memory filters as Qwen3-Next and Qwen3.5. - reuses the existing indexer, per_layer_token_embd, SSM and compress_ratios keys as-is. - the PLE table row count is read back from the file rather than recomputing the vocab padding rule. llama-model-loader gains UINT64 array support. That branch previously threw, so no existing caller changes behaviour; it is needed because the PLE hash multipliers do not fit in int32.
Implements the decode graph for Qwen3.8-Flash-Next: the hyper-connection residual stream, gated delta net layers, the MoE block with its gated shared expert, and dense full attention. The QSA indexer and the PLE n-gram embedding are not wired up yet and land in later commits. Hyper-connections are implemented here rather than shared with deepseek4.cpp. The two formulations agree on the [n_embd, hc, n_tokens] layout and little else: DeepSeek-V4 mixes with a full-rank projection and Sinkhorn-normalises it, whereas this model uses a low-rank down/silu/up sigmoid gate and collapses by a plain mean. Only the ~10 line stream mean is genuinely common, so sharing would mean touching DSV4's hot path and its three fused CUDA ops to reuse very little. What is reused is the substantive part: the LLM_KV_HYPER_CONNECTION_* keys, the n_embd_out_impl wide-residual support already in the loader, and the layout convention. Also allows a checkpoint to carry no PLE layers at all, which makes it possible to bring the graph up and validate it in stages. Validated against vLLM, the only working reference implementation. On a scaled-down model with an init scale large enough to give non-uniform logits, agreement with vLLM sits at the numerical noise floor: llama.cpp f32 against its own bf16 gives 84.3% top-1 agreement over 255 positions, and this graph against vLLM gives 85.1%. The comparison was calibrated by seeding three deliberate bugs (silu instead of sigmoid on the delta net gate, dropping the 1/hc scale in the mix, dropping the 2x in the combine); each drops top-1 to between 0% and 11%, an order of magnitude below the floor.
Adds the per-layer embedding: a custom I32 graph input hashes each token with its ngram_size-1 predecessors host-side and the result is a plain row gather over the shared table, the same shape gemma3n's per-layer embedding uses. The hash has to run on the host because the splitmix64-derived multipliers reach 2^45, so the products need 64-bit integers and an xor, neither of which ggml has. Predecessors that fall outside the ubatch come from a small per-sequence history on the model, mirroring the per-request ngram_context the reference carries. It is only trusted when contiguous with the incoming position, so a fresh prompt or a rewound cache falls back to EOS padding rather than hashing against stale tokens. The depthwise conv is written out as a sum of shifted, per-channel-scaled copies rather than through ggml_conv_1d_dw, which carries a correctness warning upstream. Verified two ways. The row indices match a transcription of the reference's tensor formulation exactly, 1024 of 1024 rows, including sequences with EOS tokens sprinkled through them to exercise the segment reset. Separately, with PLE placed on layer 0 so its input is just the token embedding, ple_embd and ple_gated_value match a PyTorch computation from the same checkpoint to every printed digit. End to end over 1023 scored positions the port sits the same distance from vLLM with PLE as without it, 6.3 points of top-1 against 6.0, so PLE costs no accuracy relative to the rest of the model. That common offset is vLLM's bf16 activations, which cannot be removed: its QSA kernel refuses float32. Two bugs found along the way, both caught by the row-index check. The history was read and updated in the same pass, so a token early in a ubatch could pick up an earlier token of that same ubatch as prior context; it is now snapshotted first. And an EOS token was cutting its own context, where the reference takes the last EOS strictly before the position, so a boundary only hides tokens from the positions after it. Known gap: the conv carries no state across ubatches, so it is exact only for a prefill that starts at position 0. Chunked prefill and decode need the conv state wired into the recurrent memory, and the conv branch itself is still numerically unverified because the fixture zeroes its weights.
The PLE depthwise conv was zero-padding on the left, which is only right for a prefill that starts at position 0. Decode and chunked prefill saw a truncated history for the first (kernel-1)*ngram_size positions of every ubatch. The PLE module sits on a layer that is also a delta-net layer, so both need a conv history in the same recurrent row. Rather than plumb a per-layer state size through build_rs and build_conv_state, the row is widened once and each convolution addresses its own slice through a local helper. n_embd_r() gains the extra span, which is zero for every other architecture because it is derived from ple_n_heads. Verified by feeding the same 1024 token sequence in chunks instead of one shot: at 64 tokens per decode the logits are bit-identical to the single-shot run, 1023 of 1023 top-1 and a maximum logprob deviation of exactly zero. At one token per decode they differ slightly, but the no-PLE model differs more under the same test (94.6% against 97.1%), so that is the usual gemv-versus- gemm accumulation difference and not the state. The conv branch is also no longer unverified. With non-zero conv weights the port sits 6.3 points of top-1 below the numerical floor, the same distance as with the weights zeroed and as the model with no PLE at all, so the branch adds no error of its own. test-llama-archs passes every existing architecture at 0.00e+00, including the delta-net models that share this code path.
build_rs writes into the state tensor in place, zeroing one row and copying the carried-over states, so calling it twice for the same layer let the second call clobber the first write-back. The PLE layer is also a delta-net layer, so that is exactly what happened: both convolutions gathered the same row. They now share a single gather per layer. The earlier claim that the conv state was carried correctly was tested on a fixture whose conv weights are zero, where the branch contributes nothing and chunking matches trivially. Re-running with non-zero conv weights showed the divergence, growing with the number of ubatch boundaries: 97.1% top-1 at one boundary down to 90.2% at seven. With the shared gather it is bit-identical to the single-shot run at every chunk size tried, 512, 128 and 64, with a maximum logprob deviation of exactly zero over 1023 positions. The delta-net-only model stays bit-identical too, so nothing regressed there. Also derive the delta-net conv channel count the way load_arch_tensors sizes wqkv instead of from ssm_d_inner. The two agree for this model, but n_embd_r() only bounds the row and the convolution has to match the tensor feeding it. test-llama-archs previously aborted on this architecture and took every later architecture with it. qwen4exp is marked MoE-only, given the hyper-connection keys and an ssm_d_inner consistent with its tensor derivation, and skipped for now: the hyper-connection keys written by get_gguf_ctx are not reaching the synthesised file, which needs a separate look. The suite completes again, 124 architectures at 0.00e+00.
Groundwork for qwen4exp's QSA sparse attention. Its indexer needs a per-token key history for the full-attention layers, but a hybrid model cannot use llama_kv_cache_dsa: that class derives from llama_memory_i rather than llama_kv_cache, and llama_memory_hybrid constructs its attention cache directly. No existing architecture pairs recurrent state with a sparse indexer, so there was nothing to reuse wholesale. llama_memory_hybrid therefore gains a third, optional cache, shaped the same way llama_kv_cache_dsa shapes its lightning-indexer cache: a copy of hparams with n_head_kv forced to 1 and n_embd_head_k_full set to indexer_head_size. It is built only when a filter_idx callback is passed, which defaults to nullptr, so every existing architecture gets exactly what it got before. The per-sequence operations and the batch preparation forward to it under a null check, matching how the DSA cache prepares its two caches over the same ubatches. test-llama-archs passes all 124 architectures at 0.00e+00, including the 12 in the hybrid family that share this code. The qwen4exp fixtures are unchanged: same logits against vLLM, and chunked evaluation still bit-identical to single-shot.
The full-attention layers of this model do not attend to everything. An
indexer scores one mean-pooled key per block of compress_ratio tokens and
keeps a budget of the best blocks, plus the tail of tokens that do not yet
form a complete block. Below indexer_top_k + compress_ratio - 1 cached
tokens every block fits in the budget, so the result is exactly dense.
What is reused rather than rebuilt:
- the mask machinery. build_attn's DSA overload already turns a list of
token indices into a KQ mask via ggml_set_rows, so that block is lifted
out verbatim into build_attn_mask_top_k and shared with a new overload
on llm_graph_input_attn_kv. DSA's node sequence is unchanged; the new
overload exists because llama_kv_cache_dsa assumes MLA and cannot be
dropped into a hybrid model.
- the indexer key cache, which is the optional third cache added to
llama_memory_hybrid in the previous commit. It holds raw keys, because
pooling happens before the norm and the rotation.
The graph expands block scores rather than block indices: giving every
token of a block its block's score needs only a gather, where expanding
indices would need an integer multiply-add that ggml has no op for. Since
the budget is a whole number of blocks and a block's members tie exactly,
the cut still lands on a block boundary.
Everything that depends on cache layout is computed host-side in
set_input_qsa. Blocks are cuts of the position line rather than of the cell
array, so nothing assumes the cache is contiguous.
Measured on the tiny fixture against vLLM, comparing the selected token
indices directly rather than the logits:
below the budget selection identical, and 1024-token logits are
bit-identical to the pre-QSA dense path
above the budget mean jaccard 0.975
The direct index comparison is what made this correct. The reference
rectifies each head's dot product before summing over heads, which an
earlier reading of it had missed; on logits alone the resulting port looked
fine, because on a randomly initialised fixture the known-correct dense
path already disagrees with vLLM by more than the bug did. Comparing the
indices showed 0.794, and fixing the ReLU moved it to 0.975.
The indexer cache found its own slots, independently of the attention cache. Both are the same size and see the same ubatches, so in a straight-through prefill they agree, which is why every fixture and every single-shot parity run passed. They drift once the context is being rewritten between turns, and then the QSA top-k indices, which are applied against the attention mask, point at the wrong cells. The seven-turn chat test caught it on the third turn: llama-server aborted on the assertion that the two caches report the same n_kv. The cache is a side buffer addressed by the attention cache's cells, so it now takes that cache's slot layout instead of computing one. Applying that layout also marks its cells identically, so the two agree cell for cell by construction rather than by coincidence, and the assertion can no longer fire. Inert where the caches already agreed: test-llama-archs green at 126 archs and 0.00e+00, and the 4096-token tiny fixture is unchanged at max logit delta 0.0.
The old note guessed that the hyper-connection keys never reach the file. They do: dumping the gguf_context handed to llama_model_init_from_user shows both among its 67 KVs, and the loader still reports one missing.
The arch was skipped with a note guessing that the hyper-connection keys
never reached the synthesised file. They did. The suite builds a model, then
saves and reloads it, and llama_model_saver did not re-emit those keys, so
the failure was in the roundtrip leg rather than the first load. Three gaps,
all in shared code and all additive:
- add_kv_from_model wrote no hyper-connection, compress-ratio or PLE keys.
The PLE group only means anything whole, so it is written or omitted
together; the rest follow the file's existing style of writing every key
unconditionally, since an architecture that does not read one is
unaffected by a zero.
- the saver had no uint64 path at all, which the PLE hash constants need.
- add_tensors_from_model enumerates model-level tensors by hand and was
missing per_layer_tok_embd and the three final-mixer tensors.
Two smaller fixes on the qwen4exp side, both found by running the test:
- build_qsa_top_k divided by the compression ratio before asserting it was
non-zero, so a file without the key crashed instead of reporting.
- a layer with no compression ratio now falls back to dense attention,
which is what the model computes below the budget anyway. The test then
has to write a ratio to reach QSA at all, and an indexer key length no
narrower than n_rot, since the indexer ropes with the main attention's
rotary width.
Full suite: 126 archs, qwen4exp at 0.00e+00 with roundtrip OK. The tiny
fixture is unchanged, max logit delta 0.0 against the pre-QSA dense run.
The n-gram table arrives as 128 shards that were held in a dict and then torch.cat-ed, so the peak was the shards plus the concatenation: around 300 GB of RSS on the real checkpoint, which rules out machines that could otherwise convert this model. Each shard is now written straight into a memory-mapped file at its final row offset and dropped, so the resident set is one shard and the rest is the page cache's problem. The temporary file sits beside the output and is removed once the write finishes, including on failure. Shards other than the last must be uniform for direct placement, which is asserted rather than assumed, and a shard arriving before the stride is known is held instead of misplaced. Verified on the tiny fixture: the resulting GGUF is byte-identical to the one the concatenating path produced (md5 2d274efac91ad1e9a6007efb0687e597).
tensor_type_fallback demotes a tensor whose ncols is not a multiple of the target's block size, but its switch only enumerates the 256-block types. A target that is already a 32-block type (iq4_nl, q4_0, q5_0, q8_0, ...) falls into default: and throws, even though the function already knows how to answer that case: the ncols check right below the switch resolves an unrepresentable shape to F16. Route those types into that check instead of throwing. Only paths that abort today change, so no quantization that currently succeeds is affected. Found on a 4-wide depthwise conv kernel. llama-quantize reported nothing but "failed to quantize model from ...", with no tensor name and no exception text, which made a quant recipe that had simply not pinned the tensor look like a corrupt model. It now names the tensor and continues.
per_layer_token_embd shares the TOKEN_EMBD category with token_embd.weight, so --token-embedding-type is returned for it before any --tensor-type pattern is consulted, and there is no way to give it a tier of its own. That grouping is fine as a default and stays the default. It is a poor fit for the size, though: on qwen4exp the table is 97.7 GiB of a 337.6 GiB BF16 file and about 46% of a 4-bit one, roughly eighty times token_embd.weight, and it is read by ggml_get_rows rather than a matmul so no imatrix ever covers it. Allow an explicit --tensor-type pattern to name it, and only it. Nothing changes unless such a pattern is passed, and token_embd.weight keeps the old precedence in either case. Measured on Qwen3.8-Flash-Next, Q4_K_M with an imatrix: the table lands at q8_0 (51.9 GiB, 113.5 GiB total) by following --token-embedding-type, and pinning it q4_1 gives 30.5 GiB for 92.1 GiB total, 19% off the file.
The per-tensor output buffer was sized `nelements * 4`, described as an upper bound. It is a very loose one: the output is at most 2 bytes per element (f16/bf16) and usually well under 1.1 (q8_0 and below), so between 2x and 4x of it is never touched. The exact size is already known here, since it is what the quantization loop writes, what new_size sums to, and what the GGUF metadata is asserted against a few lines later. On a model whose largest tensor is a few GB none of this matters. On Qwen3.8-Flash-Next it does: per_layer_token_embd is 51.2 G elements, so the buffer was 205 GB where 54 GB is needed at q8_0 and 32 GB at q4_1. Measured on that model, VmHWM of a live llama-quantize was 485 GB per process. Three of them fit in 2 TB and five did not, which is what an OOM-killed quant ladder looks like. This removes about 150 GB of that. Byte-identical output, verified against the same binary built at the parent commit: q4_K, q8_0, q5_K, q6_K and IQ4_XS, over BF16 and F32 sources, with and without a PLE table present. Six cases, six matching md5s.
The PLE row indices are computed host-side from ubatch->token, and set_input
returned early when that was null. A multimodal ubatch is exactly that case:
the mtmd layer consumes the image placeholder ids and hands llama_decode
embeddings instead. The early return left the I32 index tensor uninitialised,
so ggml_get_rows indexed a 320 M row table with whatever the buffer happened to
contain, and aborted:
GGML_ASSERT(i01 >= 0 && i01 < ne01) failed
ggml_compute_forward_get_rows
mtmd_helper_decode_image_chunk -> llama_decode
Every image request crashed. Nothing caught it because the vision work had only
ever been verified by converting an mmproj, never by running one.
The reference computes the hash over input_ids, where those positions still
hold the image placeholder, so carry that id through as qwen4exp.ple.image_token_id
and hash it. The key is optional: a file converted before it existed falls back
to the PLE EOS token, which is defined and treats the image as a segment
boundary rather than crashing.
Verified end to end with llama-mtmd-cli, a Q4_K_M base and the F16 mmproj, on a
generated image with known content. The model names the red circle, the blue
square, the inverted green triangle and reads "UNSLOTH 42", each with the right
position.
set_input_qsa asserted n_stream == 1, so llama-server could not serve this model with more than one slot unless -kvu was passed. With a non-unified cache each sequence owns its own cells, and a cell index means a different token in each stream, so a single shared mapping is wrong. - cell_blk, blk_cells and bias gain a stream dimension. At n_stream == 1 these collapse to the shapes they had, so the unified path is unchanged. - Scoring is now batched over streams. ggml_mul_mat matches ne[2] on both operands, so stream s's queries only ever meet stream s's blocks; without this sequences would score against each other's context. - set_input_qsa loops per stream and resolves cells through v_cells[seq_to_stream[seq_id]], following set_input_kq_mask_impl, instead of hardcoding v_cells[0]. - llama_kv_cache_context::get_n_stream() is added, mirroring the ns that get_k and get_v already derive from the slot info. build_attn_mask_top_k needed no change: it already expects [n_top_k, n_batch, 1, n_stream], so the top-k result is reshaped to meet it. set_input_qsa has exactly one caller, so the blast radius is qwen4exp only. Validation, UD-Q4_K_XL on one B200: - unified cache unchanged within noise: 1802.9/68.85 -> 1807.2/69.11 t/s at batch 1, 2262.5/192.43 -> 2270.1/193.75 at batch 4. - non-unified now runs at npl 1, 4, 16 where it previously aborted, and is 22% faster than the -kvu workaround at batch 16 (1205 vs 984 t/s total), since per-stream cells avoid the cross-sequence masking a unified cache pays for. - no cross-stream contamination: four concurrent sequences each carrying a distinct secret all recall their own and no other, on both cache modes. - test-llama-archs green on qwen4exp, deepseek2, gemma3n, qwen3next, llama. Note on testing: comparing concurrent output against solo output exactly is not a valid check. It failed 0/4 with no bug present, and the unified-cache control failed the same way, because batch composition changes the floating-point reduction order and near-tied tokens flip. The contamination test above is what the exit code gates on.
The QSA graph needed a build_attn that attends only to the cells named by a
top_k tensor, and the first version got it by adding a llm_graph_input_attn_kv
overload to llm_graph_context and factoring the mask construction out of the
existing MLA sparse path into a shared build_attn_mask_top_k.
That put a new arch on the shared attention path and made the deepseek32 and
glm-dsa attention build depend on a helper introduced for qwen4exp. Build the
mask in src/models/qwen4exp.cpp instead and leave llama-graph.{h,cpp} exactly as
they were: the MLA path keeps its own copy of the same node sequence.
The nodes emitted are unchanged, so this is bit-identical.
The indexer key cache was added by extending llama_memory_hybrid with an
optional third cache, and the host-side cell/block mapping that drives QSA was
added as set_input_qsa on llama_kv_cache. Both are shared classes that every
hybrid and every attention model goes through.
Move both into a new memory type, llama_memory_hybrid_idx, following
llama_kv_cache_msa: the indexer cache and the pos<->cell translation live with
the sparse-attention memory rather than in the classes that serve every other
architecture. llama-kv-cache.{h,cpp} and llama-memory-hybrid.{h,cpp} are
restored to their unmodified state.
init_batch is repeated from llama_memory_hybrid because the indexer cache has to
be handed the attention cache's slot infos, and those are not reachable through
the context the base returns. Allocating them separately lets the two caches
drift, which is what pointed QSA's top-k at the wrong cells before.
The context derives from llama_memory_hybrid_context so build_inp_mem_hybrid
keeps working unchanged, and get_n_stream is computed from the slot infos
exactly as llama_kv_cache_context did.
Behaviour is unchanged: logits over an 8192-token sequence are bit-identical to
the previous implementation, sparse and dense alike.
llama_memory_hybrid_idx forwarded clear, seq_rm, seq_cp, seq_keep, seq_add and seq_div to the indexer cache but not state_write / state_read, so a saved session dropped the indexer keys and a restored one selected QSA top-k against an empty cache. The effect is invisible until the context passes indexer_top_k + compress_ratio - 1 cells, because QSA is exactly dense below that and the indexer contents cannot change the result. The indexer section is written last rather than next to the attention cache it mirrors. As a suffix, a reader that does not expect it stops early and the trailing bytes are caught by the size check in state_load_file; placed between the attention and recurrent sections it would instead be parsed as recurrent state, which can succeed and restore silent garbage. It follows the same LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY gate as the attention cache, since a partial checkpoint deliberately skips the token-level attention caches. The indexer restores its own cells instead of taking the attention cache's restored slots. The two caches share size, padding and every sequence operation, and init_batch hands the indexer the attention cache's slot infos, so both state_read_meta calls run find_slot over identical occupancy and land on identical cells. The overrides live on llama_memory_hybrid_idx, the only memory type that owns an indexer cache, so llama_memory_hybrid and every architecture that uses it write and read exactly the bytes they did before. The session and sequence state versions are bumped because the qwen4exp state layout changed. The session path already rejects a short read via its size check, but llama_state_seq_load_file accepts one silently, so only the version check stops a pre-fix blob from being half-restored by a fixed build. (cherry picked from commit 2721542354f8e158c3217625f4e2e7b83e51e3fe)
The PLE hash of a token mixes in the ple_ngram_size - 1 tokens before it, which
a decode ubatch does not carry, so they were remembered in a map on
llama_model_qwen4exp. That is the wrong owner twice over.
A llama_model is shared by every context that loads it, and the map was keyed
only by llama_seq_id, so two contexts running the same sequence id - two server
instances on one model, or a draft/target pair - overwrote each other's window.
The next_pos guard turned that into EOS padding instead of a crash, so it
degraded quality silently.
The map was also in no state blob: grep found ple_hist in neither
llama-kv-cache.cpp nor llama-memory-*.cpp nor llama-context.cpp. A restored
context therefore failed the next_pos check on its first ubatch and hashed the
first tokens after the restore against EOS padding. This is why a session blob
round-tripped byte for byte while the restored context computed different
logits: the state was never in the bytes.
It moves to llama_memory_hybrid_idx, which is per context, is the memory type
qwen4exp always builds, and already does the per-sequence bookkeeping this
needs. Every sequence operation now carries the window with it:
seq_rm a rewind (p1 < 0) truncates the window to the surviving prefix and
moves next_pos to p0, so a rollback keeps exact context; a hole
punched in the middle leaves the window non-contiguous, so it is
dropped
seq_cp the destination inherits the source's window, truncated to the
copied position range - a copied sequence continues with the same
n-grams the source would have used
seq_keep every other sequence's window is dropped, like its cells
seq_add a shift that moves the whole window keeps it and moves next_pos with
it, which is the context-shift case; one that cuts through it drops
it
seq_div positions stop being consecutive, so an overlapping window is
dropped
clear everything is dropped
Dropping means next_pos = -1, which set_input turns into full EOS padding: the
same thing a fresh sequence gets, and the same thing this code did before it
followed the sequence operations at all, so no case is worse than before.
The state payload is a self-delimiting list, u32 count then per entry
{ i32 seq_id, i32 next_pos, u32 n_toks, i32 toks[n_toks] }, so a whole-context
save and a single-sequence save share one format and a single-sequence restore
can retarget the window at its destination seq_id. It is written after the
indexer section, last, for the same reason that one is: as a pure suffix an
older reader stops early instead of parsing these bytes as something else.
Unlike the indexer section it is not under LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY.
The window is recurrent state - it is the input the PLE convolution's own
recurrent state is derived from - and the recurrent cache beside it is written
for partial checkpoints too. Gating it would leave the server's speculative
decoding checkpoints restoring the conv state without the window that produced
it.
No further version bump: LLAMA_SESSION_VERSION 10 and LLAMA_STATE_SEQ_VERSION 3
were introduced for the indexer section in the same unreleased series, and both
changes are qwen4exp-only additions to the same blob layout.
Also fixes the padding of a short window. set_input pads a window shorter than
ngram_size - 1 up to that length, but prev() indexes the snapshot with the most
recent token last, and resize() pads at the back, so the filler EOS landed where
the immediately preceding token belongs. It now pads at the front. A window is
short at a sequence start after a one-token prefill, and after a seq_rm rewind,
which the new bookkeeping makes common.
Every architecture other than qwen4exp builds llama_memory_hybrid rather than
llama_memory_hybrid_idx, has no PLE table and never asks for a history, so
nothing about its graph, its sequence operations or its state bytes changes.
(cherry picked from commit de170364c052c68fcf63285cc0028095edb9f23c)
Rewrite the comments this series adds to the AGENTS.md rules: one or two lines, no prose hard-wrapped mid-sentence, no narrative or history, and no comment that only restates the code. Net 146 fewer comment lines, no code change. Correct the PLE image comment: mtmd does not consume the placeholder ids. An image is decoded as an embeddings-only batch, so ubatch->token is null and the per-position ids never exist here. gemma3n and gemma4 hit the same case and stand in row 0 of per_layer_token_embd; qwen4exp stands in the configured image token id instead. Read image_token_id straight from self.hparams in the converter. base.py merges text_config into the root of hparams, and the key sits at the root of config.json, so the config.json re-read was redundant. (cherry picked from commit 205840c12169057da3e8d2f65ec4ceec3e18b980)
(cherry picked from commit 4c30574f81dc1115d08078c47b6cf8c789c0a842)
(cherry picked from commit 37c8c194e6a30e4c46ac29bee3fb264f091596ef)
(cherry picked from commit 528d032b51fa3cf935ed3ef6e0fb1c7401df53b5)
f32_conv_buf held the whole dequantized tensor, which is 204.8 GB for per_layer_token_embd alone and dies with std::bad_alloc long before the work buffer is reached. Dequantize and quantize in bands of whole rows instead, capping the f32 staging at 1 GiB per band. Rows are independent and the imatrix is indexed by column, so band boundaries cannot change any output byte. Bands nest inside the existing per-expert loop so each expert slice keeps its own imatrix, and a band is kept to at least one quantization chunk per worker thread so the existing multithreading still has work. F32 sources still stage nothing and are banded by pointer arithmetic into the tensor. llama_tensor_dequantize_impl now takes a first element offset; the single caller is updated. (cherry picked from commit 658c22549613555dbce57a772be4de8509eba3ee)
… dsa-iswa header, model members, enums)
|
Good catch. Your Flash Attention diagnosis was exactly right, and your patched real-model results are enough to validate the GQA fallback itself. I pushed two follow-ups:
Local results on
I agree with your synthetic Qwen3Next diagnosis. I tried the cheap fixture-growth route, but simply increasing its head/group counts violates another delta-net shape invariant. I dropped that test-only change. The fixture needs a coherent set of larger dimensions, and it should not block the real Flash-Next fix you already validated. Please sync to
For DS4, please post the exact command, whether warm-up completes, the first short completion, and any assertion if it still fails. No need to rerun the known undersized Qwen3Next fixture right now. Performance correction: the later controlled rerun shows no general tensor-split regression on this box. Dense Qwen3.8-27B is about 1.7x faster with tensor split, while sparse Flash-Next is slower because its roughly 6B active parameters leave very small per-device expert matmuls while still paying the per-layer all-reduce cost. NUMA binding also stabilizes the four-GPU result. Recommendations should account for model workload and topology, not topology alone. |
|
Correction to my throughput note. I said tensor split was 2.5x slower and implied that was a Re-ran the historical benchmark apples-to-apples — same model (
No regression anywhere — the PR build matches or slightly beats every historical arm. And on The 4-GPU arm is bistableFour unbound runs gave two clean modes, ~12.6 and ~15.8 — not a spread. My first sample landed in The historical script binds the 2-GPU arms with
4 of 4 bound runs >= 14.68; 2 of 4 unbound at ~12.6. With GPU0/1 on NUMA 0 and GPU2/3 on NUMA 1, What is actually true about Flash-NextThe 6.20 tok/s figure holds, but it is a property of the model, not the topology:
Flash-Next has ~6B active parameters and So: tensor split is a solid win on dense models here and a loss on this sparse MoE. Sorry for the |
|
Correction noted, and thanks for rerunning the controlled matrix. I updated my previous comment so it no longer generalizes the Flash-Next result to the whole topology. The useful distinction is clear now: tensor split is a strong win for the dense 27B model on this box, while Flash-Next loses because its sparse 6B-active workload leaves tiny per-device expert matmuls but still pays the all-reduce cost. The NUMA binding result is also worth keeping with any four-GPU benchmark. No code change is needed from this correction. The FA, GQA, VRAM, and coherent-output findings remain intact. |
|
Synced to 1. Flash-Next Q2,
|
|
I added the central policy in Known-safe undersubscribed dense GQA still takes the mirrored path. For every other tensor, the splitter now checks the number of granularity-aligned units before producing device slices. If there are fewer units than devices, it stops at split computation with the tensor name, segment, unit count, device count, and the useful remedies instead of emitting a zero-width slice and letting a later operator fail or produce garbage. This should turn the undersized synthetic Qwen3Next case into one direct diagnostic rather than a later Local results on the new head:
The requested real-model checks remain Flash-Next Q2 on four devices and DS4 with the even split first. If either stops, the new first error should now be the only trace we need. |
|
Synced to 1. Flash-Next Q2,
|
| check | 163517b9a |
3042c600b |
|---|---|---|
| Flash-Next Q2, 4 dev | coherent | coherent, VRAM identical |
DS4 -ts 1,1,1,1 |
OOB write at load, tensor unidentified without instrumentation | named at source, one clear diagnostic |
DS4 -ts 1,1 (2 dev) |
not run | same refusal — 1 unit for 2 devices |
Did not re-run the undersized Qwen3Next fixture, per your note.
|
Thanks, Mark. The guard did its job, and your two-device check proves this is an architecture policy case rather than a merely undersized four-device split. I pushed
Local results on the new head:
Flash-Next Q2 on four P100s is confirmed coherent and unchanged at 13603 / 13633 / 13603 / 13603 MiB, so there is no need to repeat that run. Please sync to |
|
Synced to DS4,
|
| head | DS4 -ts 1,1 first failure |
|---|---|
c232282aa |
:730 SET_ROWS split mismatch (graph) |
163517b9a |
ggml-backend.cpp:472 OOB write during weight loading |
3042c600b |
central policy stop: attn_output_a, 1 unit for 2 devices |
1336de3bf |
:535 UNKNOWN axis on ROPE_BACK |
1336de3bf + ROPE_BACK fix |
:593 MUL_MAT MIRRORED x AXIS_2, after warm-up |
Steady progress each time. Did not run -ts 1,1,1,1 or -ts 3,4,4,1, since you gated those on
two devices passing.
Tree left clean at 1336de3bf; all instrumentation reverted. Happy to test whatever you push.
|
Thanks, Mark. Both traces identified missing Meta propagation rules, not another tensor-placement problem. I pushed
The matrix rule is deliberately narrow. It only accepts a replicated left operand with a right operand split over axis 2 or 3, where the split is outside the contraction dimensions. Local results on the exact head:
Please sync to |
|
Synced to DS4,
|
| head | DS4 -ts 1,1 first failure |
|---|---|
c232282aa |
:730 SET_ROWS split mismatch |
163517b9a |
ggml-backend.cpp:472 OOB write during weight loading |
3042c600b |
central policy stop — attn_output_a, 1 unit for 2 devices |
1336de3bf |
:535 UNKNOWN axis on ROPE_BACK |
1336de3bf + ROPE_BACK fix |
:593 MUL_MAT MIRRORED x AXIS_2 |
16f65b057 |
:597 MUL_MAT MIRRORED x AXIS_0, one matmul further, self-reported |
Did not run -ts 1,1,1,1 or -ts 3,4,4,1, per your gating. Tree clean at 16f65b057, no local
modifications.
|
Thanks, Mark. The new diagnostic made the next issue clear. I did not add a I pushed
This restores the grouped policy from the original DS4 tensor-split implementation while retaining the central zero-width guard and the Meta propagation fixes we added. Local results on the exact head:
Please sync to |
|
Synced to
|
| head | DS4 -ts 1,1 first failure |
server ready? |
|---|---|---|
163517b9a |
OOB write during weight loading | no |
3042c600b |
central policy stop — 1 unit for 2 devices | no |
1336de3bf |
:535 UNKNOWN axis on ROPE_BACK |
no |
16f65b057 |
:597 MUL_MAT MIRRORED x AXIS_0 |
no |
4b0e2ee98 |
:597 MUL_MAT AXIS_2 x AXIS_2 |
no |
4b0e2ee98 + the rule above |
ggml-backend.cpp:552 memset_tensor unimplemented |
YES, 185 s |
Tree left clean at 4b0e2ee98; all local changes reverted. Happy to test whatever comes next.
|
I go away for a week and this somehow turns into a DS4 test PR. |
That will teach you to take time off |
|
Thanks, Mark. Your tested That head also ports the upstream Meta buffer
Local results on the exact head:
Please sync to |
That's what you get for taking a vacation :P Synced to 4x P100 sm_60 @ 150 W / 1063 MHz, The one remaining stopAs pushed, the run reaches server-ready and then aborts on the first decode (observed on 2 Rather than guess a rule I instrumented the op to print its sources' split states. Every All four sources MIRRORED. So the indexer needs no rule of its own, just routing — it sits with case GGML_OP_DSV4_HC_COMB:
case GGML_OP_DSV4_HC_PRE:
- case GGML_OP_DSV4_HC_POST: {
+ case GGML_OP_DSV4_HC_POST:
+ case GGML_OP_LIGHTNING_INDEXER: {
split_state = handle_generic(src_ss, /*scalar_only =*/ true);
} break;
Results with that one line addedThree prompts per arm, not one: a fresh prompt, an unrelated prompt (forces a full sequence
The even tensor arms are exact to the megabyte; layer split at Sample completion, 4 devices,
and on the primes prompt it returns 2, 3, 5, 7, 11 with One thing I nearly reported as a bug, and the control that stopped meThe uneven It is not.
|
|
thanks for fix the gpu split issues but its still doest working well when prompt is getting to process. meaning its worked initially but after 30k context its crashed even having the space in the gpu. gopi@gopi-Super-Server:~$ echo 0 | sudo tee /proc/sys/kernel/numa_balancing |
|
Thanks, Mark. That completes the DS4 hardware matrix. I pushed Local results on the exact head:
Your patched-tree results already cover the important real-hardware cases on this exact rule: coherent output on 2 devices, 4 devices, and uneven I am tracking the long-context Qwen4Exp allocation report separately. Its log shows a 5.96 GiB compute-buffer request failing with |
5ff1540
into
TheTom:feature/turboquant-kv-cache
6 participants
Imports upstream PR ggml-org#27742 onto feature/turboquant-kv-cache using a clean sequential cherry-pick, with upstream-compat shims committed on top.
What this contains:
Verified:
Sequential cherry-pick log: /tmp/qwen4exp-sequential.log (26 commits, 0 cherry-pick failures)