Commit Graph
76 Commits
Author SHA1 Message Date
ZihaoMuandLZ-QWQ 9cacd0896c fix(moe): publish rocm cpu expert inputs before notification (#764)
* fix(moe): prevent stale cpu expert inputs in rocm graphs

* fix(moe): publish rocm cpu inputs with graph events

* fix(moe): qualify rocm kernel io with ordered copy fallback

* fix(moe): enable locked hybrid staging on rocm

---------

Co-authored-by: LZ-QWQ <lz1830837889@163.com>
2026-11-02 21:29:42 -08:00
Xiaoze Fan 3d23551c5b perf(cpu-moe): speed up the nvfp4 W4A8 kernels (#851)
The VNNI dots were compute bound per core; with fewer instructions per weight they reach the DRAM read bandwidth.
The W4A8 kernel now follows FREETOKEN_CPU_MOE_ISA.
2026-11-02 15:05:17 -08:00
Xiaoze Fan dacac5f5fa feat(spec): add MTP drafts for the Qwen3.5 family and Qwen3.8-Flash-Next (#844)
* feat(spec): add the MTP proposer for the qwen3_5 family

* test(models): read the qwen3_5 MTP head of bf16 and block-fp8 checkpoints

* feat(spec): verify qwen4_exp targets on qsa_sparse graphs with PLE verify handlers

* feat(spec): add the qwen4_exp MTP draft from the checkpoint or a Strata pack

* docs(cli): document the speculative decoding flags

* fix(spec): refuse MTP draft sources the draft cannot read before the target loads

* docs(qwen4_exp): describe the verify path in the PLE docstrings

* fix(spec): refuse a missing relative MTP draft path before the target loads

* fix(moe): read MTP head experts on the host and require both stacked tensors

* refactor(spec): tell the final prefill chunk from prompt_len instead of the scheduler's ChunkedReq

* refactor(spec): let each MTP draft family check its own weight source

* fix(qwen4_exp): size the disk PLE eager staging for decode and verify rows

* perf(kernel): read the KV-index gather's chunk count from the grid, not a constexpr

* refactor(moe): let each family name its MTP expert tensors

* refactor(spec): force the MTP draft's MoE resident in the proposer and check it after load

* fix(spec): refuse MTP with dummy weights

* refactor(spec): move HiddenRelay under spec/mtp, its only user

* refactor(spec): drop MTPDraft.config, fold its embed hooks into embed_tokens

* refactor(models): note FeatureSink's graph binding, drop its runtime_checkable

* fix(models): read quantized MTP fc projections under the quant config
2026-11-02 11:37:09 -08:00
Mukkesh Ganesh ea2916d31c feat(gguf): serve gpt-oss gguf checkpoints on gpt_oss (#838)
* feat(gguf): serve gpt-oss gguf checkpoints on gpt_oss

* fix(gguf): refuse missing gpt-oss bias tensors, honor swa pattern, scale nvfp4 before bias

* fix(gguf): honor gpt-oss yarn truncate, gate the bf16 float cast, scale nvfp4 before the down bias
2026-11-02 08:58:07 -08:00
Chen Cheng ecbbce4577 fix(moe): skip inactive gguf routes in hybrid decode (#823)
* fix(moe): skip inactive gguf routes in hybrid decode

* perf(moe): run the inactive-route memset and zeroing only in hybrid decode

Only hybrid decode hands the GGUF kernel -1 routes, so mul_mat_id takes an inactive flag
and MMQ does the ids_src1 memset and the zeroing pass only when it is set. Prefill and
fused / offload decode go back to the main-branch MMQ calls.

* perf(moe): compare expert ids unsigned in mm_ids_helper

A negative (inactive) id compared unsigned sorts after every expert and equals none, so the
helper needs no remap to INT_MAX; the remap made nvcc unroll its token loop half as far.
2026-10-28 22:28:13 -07:00
Xiaoze Fan c2c3e828e8 feat(spec): verify draft tokens on the target side (#836)
* feat(spec): add the speculative verify infrastructure

* feat(scheduler): run decode steps as speculative verify steps

* chore(engine): mark the speculative decoding config fields with one comment

* feat(server): add short aliases for the speculative decoding flags
2026-10-28 21:18:02 -07:00
Chen Cheng d2977fbc13 feat(kernel): add an iq1_m mmq kernel for moe steps above 8 tokens (#835) 2026-10-28 20:48:55 -07:00
festeraeb c94535c909 fix(attention): size the decode KV tile to the shared-memory budget (#840)
Assisted-by: Claude (Anthropic)
2026-10-28 20:05:27 -07:00
Jesús Jódar Piernas 2d93085545 feat(flint): add nvfp4 modelopt and gguf support (#797)
* feat(flint): add nvfp4 modelopt and gguf support

* refactor(flint): address final review scope requests
2026-10-24 12:37:57 -07:00
LZ_QAQ 32056cf460 fix(ple): add rocm disk wait-sync support (#776)
* fix(ple): add rocm disk wait-sync support

* fix(kernel): validate hip wait-sync support before enabling it

* refactor(ple): share decode fills and sync flag allocation
2026-10-23 15:35:47 -07:00
Xiaoze Fan fd863cce3e feat: add --embed-weights host for a pinned host embedding table (#806)
* feat: add --embed-weights host for a pinned host embedding table

* refactor: check the tie from the model config and share the op walk
2026-10-23 14:41:37 -07:00
Chen Chengandzihaomu 6505b6a005 fix(kernel): convert fp32 to e4m3 in one rounding on sm_89 (#805)
* fix(kernel): convert fp32 to e4m3 in one rounding on sm_89

* fix(kernel): avoid nvidia fp8 cast asm on hip

---------

Co-authored-by: zihaomu <zihaomu@outlook.com>
2026-10-23 14:27:34 -07:00
Mukkesh Ganesh 0c50eb9d82 perf(attention): drop redundant activation copies in glm5 dsa prefill (#795)
* perf(attention): drop redundant activation copies in glm5 dsa prefill

* fix(attention): guard the dsa sparse kernels' int32 q/o addressing
2026-10-23 12:39:33 -07:00
cherry77-cloud 507a8bfa81 fix(rotary): preserve fp32 inputs in mrope torch fallbacks (#641) 2026-10-23 00:11:06 -07:00
Manish Biswal 20d5d93132 feat(gguf): read NVFP4 token_embd rows in the vendored get_rows (#801)
* feat(gguf): read NVFP4 token_embd rows in the vendored get_rows

Signed-off-by: iamanishx <manishbiswal754@gmail.com>

* fix(tests)

Signed-off-by: iamanishx <manishbiswal754@gmail.com>

---------

Signed-off-by: iamanishx <manishbiswal754@gmail.com>
2026-10-22 20:19:24 -07:00
Mukkesh Ganesh ed359eee0a feat(gguf): serve qwen3vl and qwen3vlmoe gguf checkpoints text only (#807) 2026-10-22 18:14:03 -07:00
Xiaoze Fan 19a60a5b65 perf(moe): fuse sigmoid + selection-bias routing into a triton kernel (#804)
* perf(moe): fuse sigmoid + selection-bias routing into a triton kernel

BREAKING CHANGE: glm_moe_dsa keeps its router selection bias in fp32. An FTW converted by an older build stores it as bf16 and still loads, but routes with the rounded bias; restore it with `scripts/ftw_hotfix.py --repo nvidia/GLM-5.2-NVFP4` or reconvert.

* fix(hotfix): scope the bias restore to GLM-5.2 and drop new router comments
2026-10-22 17:48:21 -07:00
cherry77-cloud 7974c9cc8c perf(moe): skip inactive block-fp8 decode routes (#665) 2026-10-22 15:43:34 -07:00
Xiaoze Fan 609191d81f feat(gguf): take image input on Qwen and Gemma-4 GGUF checkpoints from an mmproj (#803)
* feat(gguf): load the vision tower from a llama.cpp mmproj

Qwen3.5 / Qwen3.6 / Qwen3.8 (qwen3vl_merger) and Gemma-4 26B-A4B / 31B (gemma4v) / 12B (gemma4uv) take image input from a separate mmproj file: --mmproj, else one in the model's own folder. The tower runs in bf16 (quantized mmproj files are dequantized) and its config comes from the GGUF metadata plus per-arch defaults.

* docs(gguf): describe mmproj image input without a model list

* test(gguf): drop the mmproj tests

* fix(gguf): escape the model folder when globbing for mmproj files

* fix(gguf): skip auto-discovered mmproj files the text model cannot use

* fix(gguf): serve text from a GGUF that carries vision tensors itself
2026-10-22 14:56:19 -07:00
Mukkesh Ganesh 546c5fe2e4 feat(gguf): serve DeepSeek-V4 Flash and Pro GGUF checkpoints on deepseek4 (#774)
* feat(gguf): serve DeepSeek-V4-Flash GGUF checkpoints on deepseek4

* fix(gguf): refuse reap-style per-layer expert prunes in deepseek4 ggufs

* fix(gguf): apply the deepseek4 review fixes

- route the offload experts through swiglu_clamp so the GGUF MoE path applies the limit
- map model.embed -> embed on the HF row so the bf16 exemption matches
- default a wholly missing rope.scaling block to the family yarn values with a warning;
  refuse a partly present one
2026-10-21 18:40:37 -07:00
LZ_QAQ cc4aa085b6 fix(moe): enable prefill hit-d2d on rocm (#771)
* fix(kernel): order the batch-memcpy probe against the current stream

Adapt the stream dependency and busy-stream regression from #414
(https://github.com/FlashML-org/FreeToken/pull/414). Warm the JIT
before queuing the backlog so compilation cannot mask the race.

* fix(moe): enable prefill hit-d2d on rocm
2026-10-20 23:15:17 -07:00
Xiaoze Fan c5b5b7243a feat(moe): fetch locked-layer experts through pinned staging (#742)
* feat(moe): fetch locked-layer experts through pinned staging and probe the pin limit

Under hybrid, layers whose banks the pin quota could not take keep GPU slots of their own pools and fetch part of their misses through pinned staging that CPU threads fill (--moe-locked-fetch, --moe-locked-max-fetch). --moe-cpu-layers auto without FREETOKEN_PIN_BUDGET_GB now loads the banks unpinned, pins the staging, then pins layers middle-out until CUDA refuses one and OS-locks the rest.

* fix(moe): resize the locked fetch on cache rebuild and gather staging by device address

* feat(moe): log decode moe stats per layer group with --moe-collect-stats

* perf(moe): default the locked fetch to max(2, min(4, logical cpus / 4)) copy threads

* feat(moe): take the locked fetch share from the ft bench bw profile by default

* fix(moe): fuse the locked fetch into publish and gather kernels and address review findings

* refactor(moe): gather the locked staging with fast_index_copy_multi

* refactor(moe): move the locked fetch setup out of the engine and address review findings

* fix(moe): fail the boot when the locked staging cannot be pinned and keep locked layers out of the hybrid share log
2026-10-20 14:59:05 -07:00
Mukkesh Ganesh cc56b6bdcf feat(gguf): serve GLM-5.3-Flash GGUF checkpoints on glm5-next (#721)
* feat(gguf): serve GLM-5.3-Flash GGUF checkpoints on glm5-next

Adds the llama.cpp glm5-next GGUF reader (unsloth GLM-5.3-Flash-GGUF, both
the glm5-next and the older glm5next arch spellings) on the GGUF quant
kind from #690:

- parse_gguf_config derives the text config from the GGUF KV metadata,
  with per-block attention.head_count_kv laying out the KDA/DSA hybrid,
  the trailing MTP block dropped like the HF loader drops it, and the
  MLA head widths from the *_length_mla keys.
- iter_gguf_weights maps every tensor to the glm5_next state dict and
  undoes the converter transforms: A_log stored as -exp(A_log) is
  inverted, the per-head attn_k_b/attn_v_b pair (k transposed) fuses
  into the dense kv_b_proj layout the MLA absorption reads, and the
  q|k|v conv kernels concatenate on the channel axis. The mixed-type
  fused KDA in_proj (Q8_0 q|k|v + BF16 b|f_a|g_a) rides #702 segments;
  the routed experts load as GGUF expert banks at their shipped types
  (unsloth dynamic quants: IQ3_S gate/up + IQ4_XS down, Q8_0 outlier
  layers pass through natively).
- passes quant_config to embed_tokens so the Q8_0 token_embd serves
  packed; registers Glm5NextGGUFForCausalLM and claims it in the AOT
  table.

Verified end to end against unsloth GLM-5.3-Flash-UD-IQ4_XS (5 split
shards): the full engine on a 32GB GPU serves coherent output with the
experts off host banks. Text-only, TP=1.

Part of #671.

* fix(gguf): refuse unread scale tensors in glm5-next ggufs

check_scales at config parse like the llama, qwen3 and gemma4 loaders
(review on #721): a .scale / .input_scale outside an NVFP4 weight now
fails with a clear error instead of loading silently unread.

* fix(gguf): serve the quantized tensors of unsloth dynamic glm5-next quants

unsloth's UD quants quantize tensors the AesSedAI-style files keep float:
the MLA kv_b halves (Q8_0), the mHC mixers and the indexer's wk /
weights_proj / compressor gate. kv_b_proj joins the registry's
unquantized_modules (glm_moe_dsa keeps its kv_b dense for the MLA
absorption for the same reason) and the reader dequantizes through the
vendored gguf-py reference on every float-expecting read; wq_b keeps
serving its checkpoint quant packed (attention.py keeps wk and
weights_proj dense). Regression test stores the kv_b halves Q8_0.

* fix(gguf): treat the glm5 pre-tokenizer spelling like glm4

llama-vocab.cpp maps both spellings to CHATGLM4 with ignore_merges;
unsloth writes glm4 and AesSedAI's writer glm5 for the same vocabulary.
Without the mapping a glm5 file loads but silently mis-tokenizes
(tok_compare: 22,840/153,725 vocab words differ from the HF reference;
0 everywhere with it).

* fix(gguf): size the mmq tail padding by ggml type and k

pad_rows added one zero row of k columns, which covers the bytes MMQ
reads past the last row only when k is at least half a padding unit
(512); glm5-next's KDA f_b / g_b (k = linear head dim 128, Q8_0 in
dynamic quants) are the first packed tensors below that and died at the
first mul_mat. pad_rows now takes the ggml type and returns enough whole
rows to cover mmq_tail_bytes for any (type, k), consistently in the
reader's fused(), the gguf linear and embedding params and the expert
host banks. Found serving the real unsloth UD-IQ4_XS: the dry load
never runs a forward.

* fix(gguf): take the older glm5-next writer variants in stride

Two long-tail writer variants the full check sweep found (the GSQ-RCO
lineage alone is ~57k downloads/month):

- the conv kernels with a trailing unit dim: llama.cpp's glm5-next
  loader tries that 4-D shape before the 3-D one, so trailing unit dims
  in ne are dropped on float reads (they carry no elements).
- attention.head_count_kv as the generic scalar (all blocks): wrong for
  the hybrid -- llama.cpp broadcasts it and then fails tensor creation
  -- so only the list is trusted and a scalar falls to the tensor-name
  census like the pre-list writer.

* fix(gguf): glm5-next review cleanups

- refuse instead of guessing metadata the model cannot express: the
  scalar attention.head_count_kv spelling (llama.cpp broadcasts it and
  then fails tensor creation; every public file carries the list),
  missing *_length_mla keys through need(), non-uniform
  swiglu_clamp_exp / swiglu_clamp_shexp, indexer.types with non-full
  layers, expert_gating_func other than sigmoid, and rope dims as a
  NotImplementedError instead of an assert.
- call check_scales directly; aot_models comment on its own line.
- drop the GGUF test files; coverage will be handled across families.
2026-10-19 23:20:40 -07:00
Xiaoze Fan 0db5023b60 fix(quant)!: keep nvfp4 global scales in fp32 (#739)
NVFP4 weight global scales (amax / 2688, 2e-5 to 4e-4 in the served checkpoints) sit at fp16's
normal minimum; stored as fp16, up to 97% of a checkpoint's expert globals were subnormal, with
up to 1.7e-3 relative error per row.

BREAKING CHANGE: an FTW checkpoint that stores NVFP4 global scales as fp16 is refused at load;
reconvert it with `ft checkpoint`.
2026-10-19 12:33:43 -07:00
Xiaoze Fan 4d2e2e2104 perf(linear): run nvfp4 linears on nvfp4 activations on sm_120 (#737)
* perf(linear): run nvfp4 linears on nvfp4 activations on sm_120

* fix(linear): guard the b12x nvfp4 linear against inputs it mishandles

* fix(linear): use the largest per-part global in the w4a4 nvfp4 linears
2026-10-18 23:55:07 -07:00
Xiaoze Fan e4f69d9e23 perf(moe): prefill nvfp4 moe experts on nvfp4 activations on sm_120 (#736) 2026-10-18 20:10:51 -07:00
Xiaoze Fan fd410ee616 perf(moe): pipeline the nvfp4 prefill scale loads on geforce sm_86 / sm_89 (#735)
The pointer path of the Triton prefill GEMM loads the e4m3 scales one byte per thread, too small for cp.async, so every K step waits on that load and a CTA barrier. Read them as int32 words (4 scales each) on GeForce sm_86 / sm_89 so they join the pipeline.

- SCALE_WORDS constexpr in `_prefill_nvfp4_moe_kernel` (off by default)
- Triton prefill tuning as one record per (GPU class, cc); GeForce sm_86 / sm_89 turn scale words on, with 8 warps when K >= 2048, and sm_86 moves to BLOCK_SIZE_M 64 from 100 tokens per expert
2026-10-18 14:39:56 -07:00
Xiaoze Fanandandy-yang-1 ac6a15de75 perf(moe): speed up the nvfp4 moe prefill gemm on rtx gpus (#733)
* perf(moe): speed up the nvfp4 moe prefill gemm on rtx gpus

* perf(moe): keep the byte e2m1 decode at every tile size on h100

---------

Co-authored-by: andy-yang-1 <andy_yang@berkeley.edu>
2026-10-17 17:50:59 -07:00
Xiaoze Fan 31c6ef4f77 feat(gguf): apply the weight scales of nvfp4 gguf tensors (#732) 2026-10-17 13:46:50 -07:00
Xiaoze Fan 6b8f200454 feat(gguf): serve GLM-4.5, GLM-4.6 and GLM-4.7 GGUF checkpoints (#731) 2026-10-17 12:56:19 -07:00
Xiaoze Fan 11c1adb88f perf(kernel): split triton decode attention by context length and sm count (#729)
* perf(kernel): split triton decode attention by context length and sm count

* perf(kernel): merge triton decode splits a tile at a time

- stage 2 loads SPLIT_TILE splits per step and merges them with one
  online-softmax update instead of one split per iteration
- SPLIT_TILE is the launch's split cap rounded up to a power of two,
  at most 8192 fp32 elements and 64 splits; num_warps scales with it
2026-10-16 23:56:35 -07:00
Xiaoze Fan c8559fdff6 feat(cpu-moe): serve GGUF experts on the cpu and hybrid backends through ggml-cpu (#704)
* feat(cpu-moe): serve GGUF experts on the cpu and hybrid backends through ggml-cpu

* feat(cpu-moe): run float GGUF expert rows on the cpu through ggml-cpu's vec_dot

* feat(cpu-moe): build ggml-cpu for each llama.cpp x86 variant and run the best one the CPU supports
2026-10-16 22:28:34 -07:00
Xiaoze Fan 03c2c9c35e perf(kernel): rewrite the triton nvfp4 moe prefill gemm (#723)
* perf(kernel): rewrite the triton nvfp4 moe prefill gemm

* perf(moe): tune nvfp4 prefill tiles per gpu class

GeForce runs fp32-accumulate tensor MMA at half rate, so RTX parts want smaller BLOCK_SIZE_M than H100 and their own BLOCK_N / BLOCK_KB / stages.
Tables come from tile sweeps on RTX PRO 6000, RTX 5090, RTX 4090 and RTX 3080 Ti; H100-class GPUs and AMD keep the previous tuning.
2026-10-16 22:03:59 -07:00
Xiaoze Fan 4663b2b89e feat(gguf): serve MiniMax-M2 and MiniMax-M3 GGUF checkpoints (#725) 2026-10-16 19:01:30 -07:00
Xiaoze Fan 5d39d6408a feat(gguf): serve Muse-Glimmer GGUF checkpoints (#724) 2026-10-16 18:38:43 -07:00
Xiaoze Fan 21d9578bde feat(gguf): serve Llama and Mistral llama-arch GGUF checkpoints (#722) 2026-10-16 15:50:29 -07:00
Xiaoze Fanandandy-yang-1 e8c6bf1123 feat(moe): serve experts of several formats through per-format slot pools (#700)
* feat(moe): serve experts of several formats through per-format slot pools

* refactor(cpu-moe): pass each layer's cpu format with its options

* fix(moe): read only each format's layers when loading experts of several formats

---------

Co-authored-by: andy-yang-1 <andy_yang@berkeley.edu>
2026-10-16 15:13:59 -07:00
LZ_QAQ d5eb18febd fix(kernel): correct fp4 cache scale rounding (#720)
* fix(kernel): preserve fp4 scale rounding at subnormal ties

* fix(kernel): preserve native e4m3 scale calculation
2026-10-16 13:49:17 -07:00
Xiaoze Fan 6c21dd5615 feat(gguf): serve Qwen2, Qwen3 and Qwen3 MoE GGUF checkpoints (#707) 2026-10-16 13:04:29 -07:00
Xiaoze Fan 3381b6eb0a feat(gguf): run experts whose gate_up and down mix float and quantized types (#702)
* feat(gguf): run experts whose gate_up and down mix float and quantized types

* fix(kernel): map an fp32 compute type to tl.float32 in fused_moe_kernel_triton

* fix(gguf): write float expert GEMM output in the activation dtype
2026-10-15 23:20:12 -07:00
Xiaoze Fan df8d450c28 fix(gguf): reject ggml weights past the 32-bit block offsets of MMQ / MMVQ (#699) 2026-10-15 21:25:32 -07:00
Zhang Jason 7d6490c5d6 fix(rocm): clamp QSA KV tile to the device LDS budget on RDNA (#587)
The qsa_sparse_paged_attention tiles are tuned for the GB300's 228KB shared
memory; RDNA gfx11/gfx12 have ~64KB/CU, so k+v staging overflows LDS. Query the
device's actual shared-memory limit and shrink the KV tile only when it doesn't
fit, ROCm-gated so NVIDIA tiles stay byte-identical.

The gguf ROCm-compile fixes this branch also carried are now covered upstream by
the ggml-cuda vendoring in #690 (separate DEFAULT_HIP_CFLAGS + vendors/hip.h), so
this branch reduces to the QSA clamp only.
2026-10-15 20:34:20 -07:00
Xiaoze Fanandzihaomu e5d3bd3faf feat(gguf): serve GGUF checkpoints on vendored ggml-cuda kernels (#690)
* build(kernel): vendor ggml-cuda from llama.cpp 8c2134a8 behind a tvm-ffi binding

* build(deps): vendor the gguf-py reader

* feat(quant): add the GGUF quant kind and an embedding quant method

* feat(gguf): serve Flash-Next and Gemma4 GGUF on the GGUF quant kind and drop the old Q4_0 path

* fix(kernel): build ggml-cuda for plain sm_120 by turning off the native fp4 MMA

* perf(gguf): write ggml results in the activation dtype and keep MoE intermediates in bf16

* build(kernel): compress the ggml fatbin to keep the multi-arch kernel cache small

* fix(gguf): read F32, F16 and BF16 PLE tables without ggml

* fix(moe): pad the offload slot cache for ggml MMQ tail reads

* fix(gguf): support hip builds and runtime devices

* fix(gguf): copy same-type dequantize and check for missing ggml converters

* fix(gguf): load quantized qwen4exp ple_key and ple_value

* feat(gguf): run float parts of a fused GGUF linear as bf16 segments

* fix(gguf): load quantized qwen4exp hyper-connection and gdn ba weights

* fix(moe): zero the prefill staging area after the last prefill layer

---------

Co-authored-by: zihaomu <zihaomu@outlook.com>
2026-10-15 16:33:16 -07:00
Xiaoze Fanandzihaomu 2010271266 refactor(cpu-moe): split the executor into weight formats and gpu sync backends (#670)
* refactor(cpu-moe): split the executor into weight formats and gpu sync backends

- move each expert weight format into kernels/<fmt>.cpp behind the MoeFormat
  interface in format.h; the executor keeps the threads, the work split and
  the pass 2 route sum
- pass each format its banks and integer options by name; MoeParams keeps only
  the shared shape and activation fields
- key the formats by name in one table that weight_formats() exports, replacing
  the integer format ids and the fp8_block_isa_name hook
- replace the per-(layer, batch size) flag slots, pinned IO buffers and task
  table with one doorbell word MoeTask{layer, num_tokens} and one shared pinned
  IO block of max_tokens rows
- move the GPU/CPU handshake behind GpuSync with a CUDA and a HIP backend picked
  by setup.py; HIP waits on signal memory and adds explicit batch-memop graph
  nodes where the runtime does not record the native memops under capture
- raise on graph capture when the handshake is off and host functions are not
  replay-safe
- allocate format scratch on 64-byte boundaries
- size max_tokens by the explicit cuda_graph_bs as well

Co-authored-by: zihaomu <zihaomu@outlook.com>

* perf(cpu-moe): skip the worker pool for tasks without routed experts

* perf(cpu-moe): publish and combine the cpu handshake with kernels

* fix(cpu-moe): clean up failed hip capability probes

---------

Co-authored-by: zihaomu <zihaomu@outlook.com>
2026-10-13 20:12:58 -07:00
95ac86f215 feat(cpu-moe): serve block-fp8 experts on the cpu and hybrid backends (#662)
* feat(cpu-moe): block-FP8 expert GEMV (W8A16) for the cpu/hybrid backends

Block-fp8 checkpoints (Qwen3.5/3.6-FP8, DeepSeek-V3-style, GLM) had no CPU expert
path at all:

    NotImplementedError: --moe-backend cpu/hybrid ... supports
    ['bf16', 'ds_fp4', 'mxfp4_triton', 'nvfp4', 'q4_0'] formats,
    but this checkpoint's experts are 'fp8_block'

so every expert miss had to stream over PCIe at ~25 GB/s while the CPU sat idle.

The bank is fp8-e4m3 row-major [rows, K] with one bf16 scale per 128x128 block --
K contiguous per output row, unlike mxfp4's transposed bank, so it maps straight
onto a bf16 dot product. e4m3 -> bf16 is *exact* (3 mantissa bits into 7, and
bf16's exponent range covers e4m3's whole 2^-9..448 span), so the AVX-512 path
widens weights in-register and feeds `_mm512_dpbf16_ps` against the bf16
activations: identical products to the reference's fp32 math, differing only in
summation order -- the same latitude the bf16 dot already takes.

The normal range rebases the exponent with a shift and a constant +120; exp == 0
is subnormal (m * 2^-9) and does not obey that, so those lanes blend in from a
small table. AVX2 does the same in fp32; scalar goes through the existing 256-entry
e4m3 LUT.

`ft bench bw` gains an fp8_block CPU measurement (it could already build the banks).
Measured on main with this commit alone, `--dtype fp8_block --reps 5`, 2x Xeon Gold
6526Y:

                        CPU MoE      PCIe    ratio   backend picked
    main             not measurable  25.1     --     offload
    main + this      53.7 GB/s       25.1    2.14x   hybrid

Before this commit the format reports no CPU number, so the backend picker has
nothing to weigh and selects offload. The ratio is what decides the backend, so
adding the kernel is what lets a block-fp8 deployment use the CPU at all.

End to end on Qwen/Qwen3.6-35B-A3B-FP8, median of 3 generations at matched GPU
expert residency:

    residency   offload   hybrid    delta
    10%           43.94    50.94    +15.9%
    25%           82.65    55.12    -33%

The win is real but conditional: hybrid pays off once the miss rate is high enough
that PCIe is the bottleneck, and costs throughput when enough experts are resident
that the CPU becomes the limiter. See the PR discussion -- the ratio makes
`--moe-backend auto` select hybrid for every block-fp8 checkpoint, which is right
for the oversubscribed case this project targets and wrong for a model that nearly
fits, and the bench models bandwidth but not miss rate.

Correctness is checked against the Triton block-fp8 decode kernel on identical
banks and routing (bs 1/2/5), plus an exactness check on the e4m3 -> bf16 widening
over all 256 codes. A weight-format ABI probe (`max_weight_format_id`) makes a
stale prebuilt `.so` fail loudly instead of falling through to the wrong dequant
branch and returning silently wrong numbers.

Assisted-by: Claude

* fix(cpu-moe): honor padded fp8 scale strides

The scale banks' trailing dimension is padded so per-expert rows land on
a 16-byte boundary for the fused copy (kernel/aot_models.py:
fp8_block_scale_pad), so the row stride is not ceil(K/128). Recomputing
it from the unpadded block count reads scales at the wrong offset.

Qwen3.8 is the case in the tree: models/qwen3_5_moe/weight.py records
down_scale as 20x5 bf16 = 200 B, and 200 % 16 != 0 pads cols 5 -> 6.
The benchbw workload (H=2048, I=512) and DSV4-Flash (H=7168, I=2048)
both happen to give rows*cols divisible by 8, so neither trips it.

Read the stride off the tensor and pass it through to the kernel, and
let the shape check assert the unpadded extent is present rather than an
exact shape. The fixture now allocates padded banks, so the existing GPU
cross-check covers this too; the GPU kernel was already stride-aware.

* test(cpu-moe): add float64 parity reference and an avx512f fp8 tier

The existing GPU-vs-CPU check compares two fp32 accumulations at a 2e-2
tolerance, which cannot see accumulator precision loss. Add a float64
reference over mixed-sign, full-range e4m3 weights and zero-mean
activations, and force each ISA tier explicitly so the sweep is real.

Also add an avx512f tier between avx512bf16 and avx2, so CPUs with
AVX-512 but no BF16 extension (Skylake-SP through Ice Lake-SP) stop
falling back to 8-wide fp32.

Measured on Xeon Platinum 8368 (Ice Lake-SP, no avx512_bf16), three
repetitions each: +19% single-thread cache-resident (3.31 -> 3.94 GB/s,
disjoint intervals), but only +4% under `ft bench bw --dtype fp8` at 75
threads (60.4 -> 62.7 GB/s, overlapping intervals). At that thread count
the kernel is memory-bound, so the project's own benchmark cannot
resolve the tier; the single-thread gain is the honest figure.

* test(cpu-moe): cover fp8 dispatch and cpu-only regressions

The new tier shipped without test coverage: deleting the avx512f branch
from select_fp8dot left every test green, because isa_name() reports
select_dot() (the bf16 chain) rather than the fp8 dispatch. Injecting a
ragged-tail bug into dot_fp8_block_avx2 was invisible for the same
reason -- no test ran numerics under a forced tier.

Force each available fp8 tier through the float64 parity GEMV, expose
the selected fp8 tier so the assertion can see it, and make the padded
bank test run a GEMV instead of only constructing an executor.

Also move the module-level CUDA guard onto the tests that actually need
a GPU. A CPU MoE kernel's most relevant hosts are the ones without a
usable GPU, and all ten tests were skipping there.

* review: force each ISA tier in the fp8_block test, and stop the docstring overclaiming

The comparison test only varied batch size, so all three cases ran
whatever tier this machine auto-selects and the scalar / AVX2 /
AVX-512-BF16 fp8_block kernels were never compared against the GPU at
all.

Parametrise FREETOKEN_CPU_MOE_ISA over the four tiers. Setting the
variable alone would not be enough: pick_isa() caps *downward*, so asking
for a tier the CPU or build lacks silently returns a lower one and the
test would pass four times over the same kernel. The executor latches its
choice into isa_name() at construction, so read it back and skip when it
does not match what was asked for. On a machine with all four, that is 4
x 3 cases against a 8-expert, 256-hidden problem.

test_subnormal_and_sign_decode_exactly said it validated that "every e4m3
byte must widen to the value the reference LUT gives", but it never calls
the CPU kernel -- it checks that bf16 round-trips e4m3 losslessly and
that subnormals are finite and small. That is the premise the widening
rests on, which is worth pinning, but it is not the widening. Renamed to
test_bf16_carries_every_e4m3_byte and the docstring now says what the
body asserts.

Not run here: this file is skipif(not torch.cuda.is_available()), and
this machine has no GPU and no torch. Compiles clean and ruff reports the
same two pre-existing E741 findings before and after.

Assisted-by: Claude

* fix(moe): let fp8_block experts decode on the cpu executor

* fix(cpu-moe): build the bf16 fp8 tier only when the compiler has avx512bf16

* fix(bench): version the bw profile and ignore older ones

* perf(cpu-moe): read fp8 weights as fp16 bits in the cpu dot kernels

* refactor(cpu-moe): write the fp8_block bank checks like the other formats

* refactor(cpu-moe): name the fp8 cpu symbols fp8_block

* docs(server): list fp8_block in the --moe-cpu-layers help

---------

Co-authored-by: Gabriel A. Devenyi <gdevenyi@gmail.com>
Co-authored-by: ChenyuHe <hechenyu@zju.edu.cn>
2026-10-11 22:12:41 -07:00
Chen Cheng ee9789ff6c fix(kernel): avoid nan in triton sampling softmax when a chunk starts with -inf (#654) 2026-10-10 01:17:57 -07:00
Xiaoze Fan 2d8b9cc1ea perf(kernel): avoid the activation layout conversion in fp8 moe decode and read mxfp4 gemv weights as column pairs (#653) 2026-10-09 00:15:07 -07:00
ZihaoMu 9b585b72d0 fix(rocm): gate CUDA-only optional backends on ROCm (#134)
* fix(rocm): gate cuda-only optional kernel integrations

* test(rocm): narrow optional-backend regression coverage
2026-10-08 19:55:57 -07:00
Xiaoze Fanandzihaomu 803a3b4802 fix(kernel): correct jit tensor matcher checks and pdl arch guards (#638)
* fix(kernel): correct jit matcher options and pdl arch guards

- with_dtype<Ts...>(sym) and with_device<Codes...>(sym) applied no options, so their dtype and device checks accepted anything
- index and store list kDLROCM, which the ignored check let through
- emit griddepcontrol only for !FREETOKEN_USE_ROCM && __CUDA_ARCH__ >= 900; it failed ptxas below sm_90 and does not exist on HIP
- declare the launch attribute cache only on CUDA
- use kWarpThreads for the index and store launch geometry

Co-authored-by: zihaomu <zihaomu@outlook.com>

* fix(kernel): treat zero as a matched size in tensor matcher

- a zero dimension counted as unset, so empty inputs failed with "Size value is not set" and a zero-row tensor did not catch a mismatched one

* fix(kernel): return early on empty index and store batches

- an empty batch launched a zero-sized grid, which CUDA and HIP reject; the leftover error then failed the next GPU call

Co-authored-by: zihaomu <zihaomu@outlook.com>

---------

Co-authored-by: zihaomu <zihaomu@outlook.com>
2026-10-08 13:31:44 -07:00
Xiaoze Fan 078132480e perf(kernel): skip masks and out-of-window tiles in triton extend attention (#632)
* perf(kernel): skip masks and out-of-window tiles in triton extend attention

* perf(kernel): drop the sm90-only register cap and k/v pipelining in triton extend attention
2026-10-08 00:36:55 -07:00