[None][feat] Kimi K3 on modeling_v2: stateful catalog entries, MNNVL and collective decode kernels, drafter attention, MoE route B - #19841

Draft

vsabavat wants to merge 183 commits into

NVIDIA:mainfrom

vsabavat:k3-up/k3-rest

Commits on Oct 2, 2026

[None][perf] Cut host work between speculative decoding steps
Vasanth Sabavat
committed
1c1d49d
[None][feat] KDA verify: optional per-token states in the V2 hybrid cache manager
Vasanth Sabavat
committed
f6cb23b
[None][feat] One-model speculative decoding: the worker produces the target logits
Vasanth Sabavat
committed
70ae504
[None][fix] spec_step_copies: store zeros past a row's accepted tokens
Vasanth Sabavat
committed
b44c537
[None][test] spec_step_copies tests: side stream waits for the current stream
Vasanth Sabavat
committed
92d67f3
[None][fix] KV cache manager V2: keep the SSM pool at its live floor with per-token KDA states
Vasanth Sabavat
committed
f3cf23f
[None][fix] MNNVL all-reduce: keep a grown workspace's predecessor alive for captured graphs
Vasanth Sabavat
committed
5e16d82
MNNVL all-reduce: run the workspace-growth graph test in the GB200 multi-GPU list
Vasanth Sabavat
committed
f0ff9a8
[None][chore] MNNVL all-reduce test: yapf the workspace-growth test
Vasanth Sabavat
committed
551ca16
[None][test] MNNVL all-reduce: one-rank test of the workspace's capture guard
Vasanth Sabavat
committed
df9d9bf
MNNVL all-reduce: count Lamport arrivals per CTA in the RMSNorm-fused kernels
Vasanth Sabavat
committed
fd1bddc
MNNVL all-reduce: start every CTA of the cluster before the RMSNorm cluster reduction writes
Vasanth Sabavat
committed
7a7ccf3
[None][fix] Kimi K3 attn_res_fwd online kernel: order the ws_stats reads before the next chunk's writes
Vasanth Sabavat
committed
f467211
[None][fix] Kimi K3 attn_res: cross-proxy fence before a consumer releases a TMA-filled slot
Vasanth Sabavat
committed
26cb3c3
[None][test] Kimi K3 attn_res: check the cross-proxy fence before each slot release
Vasanth Sabavat
committed
05e9e4f
[None][fix] kdaDecode legacy: __syncwarp before the lane-0 block-reduce stores
Vasanth Sabavat
committed
af3e384
DFlash: start the dummy slot at context length 0 every step
Vasanth Sabavat
committed
1e0dceb

Commits on Oct 3, 2026

Merge #19830's head af3e384175 (kdaDecode legacy: __syncwarp before the lane-0 block-reduce stores)
Vasanth Sabavat
committed
ee0dc11
[None][chore] Stack on the MNNVL RMSNorm-fused all-reduce cluster-barrier fix
Vasanth Sabavat
committed
6c2f854
[None][feat] Kimi K3 decode: KDA and MLA CuTe DSL kernels with their op tests
Vasanth Sabavat
committed
5db6a14
[None][chore] Stack on the one-model speculative-decoding worker contract (target logits)
Vasanth Sabavat
committed
071adfa
[None][chore] Stack on the DFlash dummy-slot fix
Vasanth Sabavat
committed
8e48156
[None][feat] modeling_v2: Kimi K3 MXFP4 target skeleton, sm_100, tp16 + moe tp4 x ep4
Vasanth Sabavat
committed
b8edc4a
[None][feat] Kimi K3 MLA attention: a caller-owned workspace
Vasanth Sabavat
committed
b45a946
[None][feat] modeling_v2 catalog: Kimi K3's stateful KDA and MLA decode entries
Vasanth Sabavat
committed
f51e780
[None][chore] Kimi K3 KDA / MLA decode: format with main's hooks
Vasanth Sabavat
committed
373b10b
[None][fix] modeling_v2 ssm/kda_decode: the op supports only the production flags
Vasanth Sabavat
committed
ce2b3c3
[None][test] modeling_v2 K3 KDA / MLA entries: sm_100 only, measured negative controls
Vasanth Sabavat
committed
95fbc35
[None][test] Kimi K3 KDA / MLA decode: sm_100 receipts and l0_b200 entries
Vasanth Sabavat
committed
8ce0792
[None][feat] Kimi K3 decode kernels (CuTe DSL): CTM GEMVs, decode GEMV, head GEMV, embedding
Vasanth Sabavat
committed
0ca3a9f
[None][perf] Kimi K3 attn_res decode RMSNorm: release the next kernel after the grid wait
Vasanth Sabavat
committed
a6828d2
[None][feat] Kimi K3 head GEMV: a caller-owned workspace
Vasanth Sabavat
committed
b6a8fb5
[None][feat] modeling_v2 catalog: the Kimi K3 single-GPU decode entries
Vasanth Sabavat
committed
11cefbd
[None][chore] Kimi K3 decode kernels: format with the repo's pre-commit hooks
Vasanth Sabavat
committed
c69ed9c
[None][test] Kimi K3 MLA op tests: drop the batch-1 identity cases
Vasanth Sabavat
committed
2700b9d
[None][feat] modeling_v2 Kimi K3 target: classify each step for the decode kernels
Vasanth Sabavat
committed
fe73486
[None][feat] modeling_v2 Kimi K3 target: its own text model, on the generic path
Vasanth Sabavat
committed
abb2564
[None][test] modeling_v2 attn_res entries: replay once before poisoning the captured outputs
Vasanth Sabavat
committed
9bc4dae
[None][test] modeling_v2 claims: a target's generic path declares every stock import
Vasanth Sabavat
committed
8bfb169
Merge U4 G3's published head 2700b9d4c2 (Kimi K3 KDA / MLA decode kernels and their catalog entries)
Vasanth Sabavat
committed
54e84c7
[None][test] modeling_v2 Kimi K3 decode entries: sm_100 receipts
Vasanth Sabavat
committed
b280442
[None][feat] modeling_v2: Kimi K3 MXFP4 target skeleton, sm_100, tp16 + moe tp16 x ep1
Vasanth Sabavat
committed
e9b3595
Merge U4 G2's head b28044256d (Kimi K3 single-GPU decode kernels and their catalog entries)
Vasanth Sabavat
committed
f2a60d5
[None][feat] modeling_v2 Kimi K3 target: KDA and MLA attention on the decode kernels
Vasanth Sabavat
committed
75d8a59
[None][feat] modeling_v2 Kimi K3 target: decode GEMVs, LM head and embedding on the G2 entries
Vasanth Sabavat
committed
0bf72e4
[None][feat] MNNVL all-reduce: Kimi K3 attention-residual epilogue (trtllm::mnnvl_allreduce_attn_res)
Vasanth Sabavat
committed
fabb6e6
[None][feat] modeling_v2 catalog: stateful entries; comm/mnnvl_allreduce_attn_res over a caller-owned MnnvlWorkspace
Vasanth Sabavat
committed
494b1d9
[None][feat] MNNVLAllReduce.allreduce_attn_res_rmsnorm, and the op's 4-rank test
Vasanth Sabavat
committed
b97811a
[None][doc] comm/mnnvl_allreduce_attn_res: every snapshot count 0-11 is certified
Vasanth Sabavat
committed
c922c7a
[None][test] mnnvl_allreduce_attn_res matrix: run on sm_100 only
Vasanth Sabavat
committed
bf1f60a
[None][feat] modeling_v2 Kimi K3 target: decode GEMV sites for the fused projections
Vasanth Sabavat
committed
67f82d4
Merge tokens' k3-up/u5-route-a-c2 75d8a597ec (KDA and MLA attention on the decode kernels)
Vasanth Sabavat
committed
376aea8
[None][fix] modeling_v2 Kimi K3 target: build the decode GEMVs' state once
Vasanth Sabavat
committed
754d706
[None][fix] MnnvlWorkspace.create: the ranks agree before allocating; the contract states the failure model
Vasanth Sabavat
committed
8c66821
[None][feat] modeling_v2 Kimi K3 target: layer 0's dense MLP on the decode kernels
Vasanth Sabavat
committed
a964500
[None][feat] modeling_v2 Kimi K3 target: the attention's projections on the decode GEMV sites
Vasanth Sabavat
committed
f224355
[None][fix] MnnvlWorkspace.create frees its communicator split when it raises
Vasanth Sabavat
committed
65cfd47
[None][fix] modeling_v2 Kimi K3 target: no measurement date in a copied comment
Vasanth Sabavat
committed
2f444c9
[None][fix] modeling_v2 Kimi K3 target: register trtllm::kda_mtp_decode on import
Vasanth Sabavat
committed
549ffd4
[None][chore] modeling_v2 Kimi K3 target: trim the text model to the target's settings
Vasanth Sabavat
committed
f01a46b
[None][fix] MnnvlWorkspace.create involves only the TP group's ranks
Vasanth Sabavat
committed
9926793
Merge #19839's head 549ffd4198 (Kimi K3 route A target: the target-owned text model and its decode kernels)
Vasanth Sabavat
committed
1af43c6
[None][chore] Stack on the Kimi K3 stateless decode entries (G2)
Vasanth Sabavat
committed
f39b747
[None][feat] MNNVL: one-shot size per call, early PDL trigger; split all-gather
Vasanth Sabavat
committed
3691863
[None][feat] Kimi K3 decode: collective CuTe DSL kernels with their op tests
Vasanth Sabavat
committed
65def52
[None][test] Kimi K3 MoE front and tail sandwich tests: torch references
Vasanth Sabavat
committed
5181297
[None][refactor] Kimi K3 collectives: caller-owned state objects
Vasanth Sabavat
committed
6fbd618
[None][fix] Kimi K3 MoE wide state: refuse graph capture; flag word docs
Vasanth Sabavat
committed
1db2a80
[None][feat] modeling_v2 catalog: Kimi K3 collective and MoE entries
Vasanth Sabavat
committed
bdad277
[None][test] Kimi K3 catalog entries: index rows and CI lists
Vasanth Sabavat
committed
20d2953
[None][test] Kimi K3 MNNVL kernel test: the split all-gather, one-shot reference
Vasanth Sabavat
committed
84fc18f
[None][test] Kimi K3 kernel lints: the MoE front, k3_moe and sandwich kernels
Vasanth Sabavat
committed
7a16d03
[None][test] Kimi K3 collective tests: names codespell accepts, lines under 120
Vasanth Sabavat
committed
a35362e
[None][chore] Kimi K3 collectives: format with the repo's pre-commit hooks
Vasanth Sabavat
committed
22629db
[None][fix] MNNVL: buffer_flags declared mutable; the split all-gather takes its world size
Vasanth Sabavat
committed
1de6f83
[None][fix] Kimi K3 collective state: the ranks agree before allocating
Vasanth Sabavat
committed
823538a
[None][refactor] Kimi K3 MoE: trtllm::k3_moe, a torch op over the caller's state
Vasanth Sabavat
committed
bd4a407
[None][test] MNNVL catalog matrices: calls queued without host sync; create() refused on every rank
Vasanth Sabavat
committed
26a3dbc
[None][test] Kimi K3 sandwich and latent exchange matrices: create() refused on every rank; ranks a step apart
Vasanth Sabavat
committed
e22949b
[None][feat] modeling_v2 catalog: moe/k3_moe over trtllm::k3_moe; moe/k3_route_quant; the front's publish flag
Vasanth Sabavat
committed
344d340
[None][test] Kimi K3 MoE front: the head geometry each TP size runs, without a GPU
Vasanth Sabavat
committed
dcb4d88
[None][chore] MNNVL split all-gather fake: format with the repo's hooks
Vasanth Sabavat
committed
d0c5e75
[None][test] Kimi K3 MoE front test: a latent code may differ only at a reference rounding tie
Vasanth Sabavat
committed
f8a29e3
[None][fix] Kimi K3 collectives: review nits
Vasanth Sabavat
committed
8c0431e
[None][fix] Kimi K3 collective state: create() frees its communicator split when it raises
Vasanth Sabavat
committed
9fbc51f
[None][chore] Stack on the modeling_v2 catalog conventions and comm/mnnvl_allreduce_attn_res (G1)
Vasanth Sabavat
committed
b30c349
[None][feat] Kimi K3 DSpark drafter attention (trtllm::k3_drafter_attn, k3_drafter_attn_qknorm)
Vasanth Sabavat
committed
4ca4bfe
[None][feat] modeling_v2 catalog: attention/k3_drafter_attn, k3_drafter_attn_qknorm; sm_100 cells for fused_qk_norm_rope
Vasanth Sabavat
committed
a6a8cb7
[None][fix] Kimi K3 collective state: create() involves its TP group's ranks only
Vasanth Sabavat
committed
90638fd
[None][doc] MNNVL catalog contracts: MnnvlWorkspace.create involves the TP group's ranks only
Vasanth Sabavat
committed
90158d9
[None][feat] modeling_v2 Kimi K3 target: attention outputs for fused post-attention steps
Vasanth Sabavat
committed
068241f
[None][feat] modeling_v2 Kimi K3 tp16_moetp16ep1: tp16_moetp4ep4's modules, copied
Vasanth Sabavat
committed
10f4a6e
[None][chore] Merge G1 9926793012: modeling_v2 catalog conventions for stateful entries and comm/mnnvl_allreduce_attn_res
Vasanth Sabavat
committed
ae282b3
[None][chore] Merge G6 a6a8cb73a3: Kimi K3 DSpark drafter attention and its modeling_v2 catalog entries
Vasanth Sabavat
committed
120913c
[None][fix] modeling_v2 Kimi K3 target: route the MXFP4 checkpoint as W4A16_MXFP4
Vasanth Sabavat
committed
03df8ee
Merge #19839's follow-ups f01a46b67f and 03df8ee849 (the trim; the MXFP4 checkpoint routed as W4A16_MXFP4)
Vasanth Sabavat
committed
35eccac
[None][chore] modeling_v2 Kimi K3 tp16_moetp16ep1: carry the trim and the W4A16_MXFP4 fix into the copy
Vasanth Sabavat
committed
fa24aca
[None][feat] modeling_v2 Kimi K3 target: KDA attention outputs on every path
Vasanth Sabavat
committed
de63263
[None][test] Kimi K3 latent exchange: the one-process test checks the TP-size refusal
Vasanth Sabavat
committed
cce3f46
[None][test] Kimi K3 wide MoE test: a seeded skewed routing instead of a router-dump replay
Vasanth Sabavat
committed
0dcbdaf
[None][chore] Merge u5b fa24aca986: Kimi K3 route B (tp16_moetp16ep1), the copy design
Vasanth Sabavat
committed
a2d912b
[None][test] Kimi K3 modeling_v2 target: run the decode-step test in the CPU stage
Vasanth Sabavat
committed
ba296fc
[None][doc] modeling_v2 catalog: 18 entries carry sm_100 receipts
Vasanth Sabavat
committed
df6da9d
[None][chore] Merge u5-route-a-next df6da9d9b0: #19839's follow-ups (l0_cpu decode-step entry, sm_100 count)
Vasanth Sabavat
committed
10eaa93
[None][test] modeling_v2 Kimi K3 collective and MoE entries: sm_100 receipts
Vasanth Sabavat
committed
db8ca58
[None][chore] Merge G4 db8ca58bd0: Kimi K3 collective decode kernels and their modeling_v2 catalog entries
Vasanth Sabavat
committed
5c37c10
[None][feat] Kimi K3 MoE: trtllm::k3_moe_m1 / k3_moe_m2, the routed experts of one or two decode tokens
Vasanth Sabavat
committed
37a8c92
[None][feat] Kimi K3 MoE: k3_moe's push build into the latent exchange
Vasanth Sabavat
committed
45cbb24
[None][test] Kimi K3 MoE: the push builds at 4 ranks against MNNVLAllReduce
Vasanth Sabavat
committed
81a0235
[None][feat] modeling_v2 catalog: moe/k3_moe_m1 and moe/k3_moe_m2 over caller-owned states
Vasanth Sabavat
committed
82fc8ad
[None][test] Kimi K3 kernel lints: the k3_moe_m1 and k3_moe_m2 kernels
Vasanth Sabavat
committed
a04ec2e
[None][feat] modeling_v2 catalog: moe/k3_moe's push form
Vasanth Sabavat
committed
f0fd94c
[None][chore] Merge RB f0fd94cfbd: Kimi K3 route B MoE engines (k3_moe_m1 / k3_moe_m2), k3_moe's push build and their catalog entries
Vasanth Sabavat
committed
9f7ad02
[None][fix] modeling_v2 Kimi K3 target: declare the long GEMV and SiTU ops it calls
Vasanth Sabavat
committed
ebfd380
[None][chore] Merge u5-route-a ebfd38001c: #19839's declared-ops fix (k3_ctm_gemv_long, k3_situ_mul)
Vasanth Sabavat
committed
85c2399
[None][fix] modeling_v2 Kimi K3 tp16_moetp16ep1: declare the long GEMV and SiTU ops it calls
Vasanth Sabavat
committed
f32084b
[None][fix] DSpark GQA drafter: keep the checkpoint's trained mask-token embedding
Vasanth Sabavat
committed
4bf2378
[None][doc] Kimi K3 collective contracts: the 16-rank runs passed
Vasanth Sabavat
committed
05ab11d
[None][chore] Merge G4 05ab11d9e8: the collective contracts record their 16-rank runs
Vasanth Sabavat
committed
9dd1f27
[None][doc] Kimi K3 MoE push form: the 16-rank runs passed
Vasanth Sabavat
committed
a55d13f
[None][fix] Kimi K3 k3_moe: compile on routing that requires grad
Vasanth Sabavat
committed
83924f8
[None][chore] Merge G4 83924f8281: k3_moe compiles on routing that requires grad
Vasanth Sabavat
committed
d017fc9
[None][chore] Merge RB a55d13f0fc: the push form's 16-rank runs passed
Vasanth Sabavat
committed
05e14ea
[None][chore] Merge u5-route-a-c2 de6326362e: the Kimi K3 target's attention outputs for fused post-attention steps
Vasanth Sabavat
committed
8ae58f7
[None][feat] modeling_v2 Kimi K3 target: a decode step's attention-residual epilogues on the fused kernels
Vasanth Sabavat
committed
eefd225
[None][feat] modeling_v2 Kimi K3 target: the post-attention all-reduce with the residual update in its epilogue
Vasanth Sabavat
committed
93f0ad3
[None][feat] modeling_v2 Kimi K3 target: the post-attention step of a decode step as one sandwich kernel
Vasanth Sabavat
committed
d1f297a
[None][feat] modeling_v2 Kimi K3 target: the decode path's one-shot ceilings for the stock MNNVL all-reduces
Vasanth Sabavat
committed
3ad8c5a
[None][feat] modeling_v2 Kimi K3 target: the MoE layers of a decode step on the K3 MoE kernels
Vasanth Sabavat
committed
d99cbd6
[None][test] modeling_v2 Kimi K3 target: the decode path's collectives at 4 ranks
Vasanth Sabavat
committed
1c9c287
[None][chore] modeling_v2 Kimi K3 tp16_moetp16ep1: carry C3 into the copy
Vasanth Sabavat
committed
ba1db7e
[None][fix] modeling_v2 Kimi K3 target: the MoE decode path hands its kernels the routing bias detached
Vasanth Sabavat
committed
ba53c84
[None][feat] SpecDecOneEngineForCausalLM: build the drafter through an overridable method
Vasanth Sabavat
committed
7dc60cc
[None][feat] modeling_v2 Kimi K3 target: decode GEMV sites for the DSpark drafter
Vasanth Sabavat
committed
31d76e0
[None][feat] modeling_v2 Kimi K3 target: its DSpark drafter on the drafter entries
Vasanth Sabavat
committed
a867199
[None][test] modeling_v2 Kimi K3 target: the DSpark drafter against the stock block forward
Vasanth Sabavat
committed
b79745b
[None][test] Kimi K3 drafter attention: the R x 7 splits
Vasanth Sabavat
committed
592ea1a
[None][test] Kimi K3 drafter attention tests: name the splits they run
Vasanth Sabavat
committed
31b6275
[None][doc] modeling_v2 catalog: the Kimi K3 drafter attention certifies R x 7
Vasanth Sabavat
committed
18f1edd
[None][feat] modeling_v2 Kimi K3 target: the DSpark drafter takes R x 7 blocks
Vasanth Sabavat
committed
0d16d09
[None][test] Kimi K3 kernel lints: the KDA, MLA and drafter kernels' rows
Vasanth Sabavat
committed
706cf2d
[None][test] Kimi K3 SiTU MoE: the TP16 shard (192 values, padded to 256)
Vasanth Sabavat
committed
f0e7d4c
[None][test] modeling_v2 MNNVL op matrices: run on sm_100 only
Vasanth Sabavat
committed
87c7e70
[None][test] Kimi K3 MLA decode view test: list it in l0_cpu, not l0_b200
Vasanth Sabavat
committed
52ca9fa
[None][test] modeling_v2 Kimi K3 construction and drift tests: list them in l0_b200
Vasanth Sabavat
committed
543ebb3
[None][test] modeling_v2 Kimi K3 routing and claims tests: list them in l0_b200
Vasanth Sabavat
committed
871ef64
[None][chore] modeling_v2 Kimi K3 tp16_moetp16ep1: carry C4 into the copy
Vasanth Sabavat
committed
8806ae3
[None][feat] Kimi K3 DSpark decode: k3_spec_accept, k3_ctx_kv and k3_markov kernels
Vasanth Sabavat
committed
32a581a
[None][doc] k3_markov: state the co-residency assumption
Vasanth Sabavat
committed
aaef4b7
[None][feat] DFlash / DSpark worker: Kimi K3 decode kernels behind k3_decode
Vasanth Sabavat
committed
a1d336d
[None][feat] modeling_v2 Kimi K3 tp16_moetp4ep4: the speculative worker's decode kernels
Vasanth Sabavat
committed
e85d7a6
[None][doc] modeling_v2 Kimi K3 tp16_moetp4ep4: the spec-worker gate's preconditions
Vasanth Sabavat
committed
4166788
[None][chore] modeling_v2 Kimi K3 tp16_moetp16ep1: carry the speculative worker gate into the copy
Vasanth Sabavat
committed
2b465f6
[None][feat] DFlash metadata: capture_view, a captured layer's slot of the capture buffer
Vasanth Sabavat
committed
32d419b
[None][perf] DFlash: the draft block's rows as a view where its slots are one run
Vasanth Sabavat
committed
1447397
[None][test] Kimi K3 DSpark decode kernel tests: list them in l0_b200 and l0_gb200_multi_gpus
Vasanth Sabavat
committed
33d3586
[None][test] Kimi K3 Markov chain test: run its checks under pytest on 4 ranks
Vasanth Sabavat
committed
5c193ac
[None][test] Kimi K3 sharded k3_spec_accept test: run under pytest on 4 local MPI ranks
Vasanth Sabavat
committed
490f289
[None][perf] modeling_v2 Kimi K3 tp16_moetp4ep4: split the DSpark drafter's fc over the TP group
Vasanth Sabavat
committed
f85b709
[None][perf] modeling_v2 Kimi K3 tp16_moetp4ep4: the DSpark drafter's residual adds and RMSNorms in its all-reduces
Vasanth Sabavat
committed
135ab0c
[None][perf] DFlash / Kimi K3 DSpark drafter: the gen requests' page-table rows as a view
Vasanth Sabavat
committed
00e4394
[None][doc] modeling_v2 Kimi K3 tp16_moetp4ep4: the drafter gate requires only its comm
Vasanth Sabavat
committed
ac6e287
[None][test] k3_ctx_kv test: drop the base-port identity cases, zero the Python path's dummy query
Vasanth Sabavat
committed
147b87d
[None][chore] modeling_v2 Kimi K3 tp16_moetp16ep1: carry the DSpark drafter's fc split and fused all-reduces into the copy
Vasanth Sabavat
committed
b7ea3f6
[None][test] Kimi K3 kernel tests: make their pool workers' functions picklable
Vasanth Sabavat
committed
19ef779
[None][perf] modeling_v2 Kimi K3 tp16_moetp4ep4: the DSpark tap from the sandwich tail
Vasanth Sabavat
committed
428f6f2
[None][chore] modeling_v2 Kimi K3 tp16_moetp16ep1: carry the sandwich-tail tap into the copy
Vasanth Sabavat
committed
fe9eaf5
[None][perf] modeling_v2 Kimi K3 target: the decode MoE's latent all-reduce as the routed experts' push form
Vasanth Sabavat
committed
4bf3512
[None][test] Kimi K3 k3_moe push test: route A's expert layout
Vasanth Sabavat
committed
41b2245
[None][chore] modeling_v2 Kimi K3 tp16_moetp16ep1: carry the latent push into the copy
Vasanth Sabavat
committed
aed1c94
[None][feat] modeling_v2 Kimi K3 tp16_moetp16ep1: the MoE decode path on route B's engines
Vasanth Sabavat
committed
d6613a1
[None][perf] Kimi K3 generic MoE path: the fused route + MXFP8 quant on k3_route_quant
Vasanth Sabavat
committed
3656bcd
[None][perf] Kimi K3 targets: route the generic MoE outside the TRTLLM-Gen kernel
Vasanth Sabavat
committed
74839d1
[None][test] Kimi K3 decode-comm matrix: route A's shared-expert slice at every world
Vasanth Sabavat
committed
4657c20
[None][fix] Kimi K3 decode MoE warm-up: no back-to-back k3_moe calls on one layer
Vasanth Sabavat
committed
0da8a02

Commits on Oct 4, 2026

[None][doc] Kimi K3 sandwich: polled input and fold store before the final grid wait
Vasanth Sabavat
committed
459e446
[None][test] Kimi K3 decode-comm matrix: free a graph's outputs before the next capture
Vasanth Sabavat
committed
c48e85e
[None][fix] DSpark K3 Markov path: release the step's acceptance and chain outputs
Vasanth Sabavat
committed
2f25c1d
[None][test] Kimi K3 DSpark drafter test: give the one-rank drafter's o_proj its TP all-reduce
Vasanth Sabavat
committed
3411d48
[None][test] Kimi K3 k3_ctx_kv: back-to-back calls share the arrival counter
Vasanth Sabavat
committed
b3b4e27
[None][test] Kimi K3 k3_spec_accept and k3_markov: arrival counters at rest after back-to-back calls
Vasanth Sabavat
committed
ae6e9a9
[None][test] MNNVL catalog matrices: armed checks wait for every rank
Vasanth Sabavat
committed
b521431
[None][doc] Kimi K3 k3_spec_accept and k3_ctx_kv: calls on a device are ordered on one stream
Vasanth Sabavat
committed
13b26f4