Repository navigation
Releases
Releases · cloudrift-ai/emmy
v0.3.27
What's Changed
- List the GLM-5.2 NVFP4 and FP8 recipes as runnable in the recipe catalog and bump Emmy to 0.3.27 by @trifonova-n in #1440
Full Changelog: v0.3.26...v0.3.27
v0.3.26
What's Changed
- Read the cache once per verify step in native serving, and measure the Qwen3.8-27B RTX 5090 golden again by @Slonegg in #1392
- Make Kernel IR the whole kernel, ahead of the PTX compiler by @Slonegg in #1395
- Trace a wide Gated DeltaNet chunk's outputs after the scan, so a 256-row program no longer asks for a 73 GiB workspace by @Slonegg in #1398
- Remove the AI pull-request review workflow by @Slonegg in #1403
- Answer a body's name questions once in Rust and trim the statement interface to what callers use by @Slonegg in #1402
- Exit early and reuse what normalization already knows, keeping every normal form by @Slonegg in #1404
- Name every kernel by its exact identity and address it by that name alone by @Slonegg in #1400
- Measure where Qwen3.8-Flash-Next's V100 gap to the async fork sits, and rank the next rounds by @6erun in #1406
- Share value numbering's parameters instead of copying them, keeping every number by @Slonegg in #1407
- Record the 256-row prefill targets of Qwen3.8-27B-NVFP4 on an RTX 5090, and fix the lowering crash they hit by @6erun in #1397
- Build each candidate cut faster by reusing unchanged terms and lifting each loop nest once by @Slonegg in #1409
- Run a Qwen3.8-Flash-Next prefill step's small routed experts 32 at a time and combine the routed rows in one pass by @6erun in #1408
- Compile Kernel IR to PTX with a Rust compiler in place of nvcc by @Slonegg in #1399
- Move the Tile IR term into a Rust crate, emmy-pure, and stop the statement-wrapper table rescanning on every insert by @Slonegg in #1410
- Run a Fold term's rewrite and recipe matching in emmy-pure, and memoize lowering on the term by @Slonegg in #1412
- Document how the classic schedule compares with the best known hand-written kernels by @Slonegg in #1416
- Record the Gemma 4 12B one-token decode rows on the RTX 5090 and refit both priors by @Slonegg in #1415
- Run FP8 products through the block-scaled instruction on sm_120, and report hand pins that did nothing by @Slonegg in #1418
- Export Qwen3.8-27B-NVFP4's native pack with 256-row prefill chunks, and fix three faults that blocked the export by @6erun in #1413
- Run the fixed-slot expert program at the decode step's width on Qwen3.8-Flash-Next V100 by @6erun in #1417
- Fix the GPU test failures on the A100, H100 and RTX 5090 by @Slonegg in #1414
- Add a hardware golden for the RTX PRO 6000 Blackwell Workstation Edition by @trifonova-n in #1422
- Run static-FP8 checkpoints' linears as W8A8, store only a rolled scan's last state, and refresh every golden by @Slonegg in #1411
- Make the PTX compiler's NVFP4 pair decode and Volta slab store as fast as nvcc's build by @6erun in #1419
- Settle the normal form of a difference of two reads, and add reference kernel times to the kernel comparison by @Slonegg in #1420
- Spell a carried state's storage as a STAGE site of the classic schedule by @Slonegg in #1421
- Write goldens so a diff lands on the node, statement or knob that changed by @Slonegg in #1424
- Re-measure the Qwen3.8-Flash-Next V100 fixed-slot expert rows on the fixed PTX compiler by @6erun in #1426
- Make make setup rebuild the Rust extensions and fetch the LFS freeze by @Slonegg in #1427
- Compose the steps of an affine recurrence as matrix products, and offer it in the carried state's fork by @Slonegg in #1425
- Add candidate files: measured schedules that are not the best known one, for each hardware-golden group by @trifonova-n in #1405
- Replay the decode graph for the last token of a prompt on Qwen3.8-Flash-Next by @6erun in #1428
- Run the nightly refit's placement check and measured-pool build on every core by @Slonegg in #1430
- Let trace read a golden's stored program, so one card's programs can be added to another card's golden by @Slonegg in #1431
- Rebuild the candidate files onto the hardware goldens' current programs, fixing main's candidate test by @Slonegg in #1432
- Record the W8A8 kernel roots of Qwen3.8-27B-NVFP4 on the RTX 5090 and time its 256-row native pack against stock vLLM by @6erun in #1423
- Cut a Gated DeltaNet readout's products once, and record its kernel sets again on three cards by @Slonegg in #1429
- Tune Qwen3.8-Flash-Next's short prefill steps on V100: symbolic pre/post rows and non-blocking expert uploads by @6erun in #1435
- Add a high-throughput GLM-5.2 NVFP4 recipe for 8x RTX PRO 6000 and bump Emmy to 0.3.26 by @trifonova-n in #1434
Full Changelog: v0.3.25...v0.3.26
v0.3.25
What's Changed
- Name the seams a cut piece offers when a piece-scoped PLACE pin names none by @Slonegg in #1289
- Let a pinned cut's remainder take its own fork, and let a bare PLACE=cut cut the whole remainder chain by @Slonegg in #1303
- Compute a cut arm's piece facts on the worker that builds the arm by @Slonegg in #1293
- Featurize a kernel set's work so the placement prior can price a recompute by @Slonegg in #1294
- Compute NVFP4 pair-table lookups in CUDA and load packed e2m1 runs as one word by @6erun in #1298
- Plan native serving to parity with vLLM, and retire the executed tuning plan by @6erun in #1307
- Re-record both RTX 5090 recipe goldens after the tuning plan's compiler fixes by @Slonegg in #1309
- Serve a recipe's native lane from an image that carries its pack by @Slonegg in #1302
- Promote both RTX 5090 recipe goldens into the hardware golden and refit the priors by @Slonegg in #1310
- Evict the instruction caches before each cold timed launch by @6erun in #1299
- Bring the Qwen3.8-Flash-Next plan up to date and state the next work by @6erun in #1312
- Measure native, plugin and stock vLLM side by side on Qwen3.6 and the 27B by @6erun in #1311
- Let the prefill chunk that ends the prompt generate the first token by @6erun in #1313
- Record milestone 4 as landed and leave its rows and gate to the card by @6erun in #1314
- Record Qwen3.8-Flash-Next's step 4 gate and first done-gate run on the V100 by @6erun in #1316
- Refuse nested tile and stage stamps the kernel builder drops by @6erun in #1315
- Add the plugin's tiny GDN hybrid to the serving regen, and split the regen into capture and GPU-free completion by @Slonegg in #1290
- Fit Qwen3.8-Flash-Next's V100 serving in the 3 GiB per card margin, and drop two decode copies per layer by @6erun in #1317
- Drop a cut piece nothing reads before fusing or joining the pieces by @6erun in #1319
- Give the rotary cache store a grid over the rows and heads it writes by @6erun in #1318
- Fold native's causal attention on tensor cores: the probabilities, the position mask and the FP8 value by @Slonegg in #1321
- Slice Qwen3.8-Flash-Next's shared expert across ranks on the V100 by @6erun in #1320
- Run the NVFP4 output head W4A4, spelled as the trunk's linears by @6erun in #1322
- Take the strict correctness rule's rounding step from the ops that round by @6erun in #1325
- Split Qwen3.8-Flash-Next's V100 pre8 down rows across blocks by @6erun in #1326
- Split native's causal attention across the keys on the chunk tier by @Slonegg in #1324
- Fix the sixteen GPU tests main fails on the dev-box RTX 5090: the stale sm_120 serving golden and two test assumptions by @Slonegg in #1327
- Give the statement module a small written interface: rendering and the body op move out, one rewrite mechanism, one entry point by @Slonegg in #1328
- Answer statement kind checks natively: 13% less body normalization, no kernel moves by @Slonegg in #1330
- Normalize a body without rounds that only check: steps that feed each other run to their own fixed points by @Slonegg in #1331
- Normalize a body in one pass: the canonical order is final on its own output by @Slonegg in #1332
- Fix what the full suite found on the one-pass normalization stack: a dead reduction under a shared name, and a timing test on a shared GPU by @Slonegg in #1333
- Reduce a Kernel IR statement kind to a name and typed fields by @Slonegg in #1334
- Declare each Kernel IR statement kind's fields and compute its statement protocol from them by @Slonegg in #1335
- Gate native serving's milestones 3, 4 and 5 on an x16 RTX 5090: 3 and 4 met, 5's long-window attention misses by @6erun in #1345
- Let a class that is not a dataclass take the wire form and the structural form by @Slonegg in #1336
- Make a statement body a plain sequence instead of a tuple by @Slonegg in #1337
- Move fusion's body assembly out of the statement module by @Slonegg in #1338
- Read a statement through its protocol, not its dataclass fields, in the clustered identity and the CPU seam walk by @Slonegg in #1339
- Give an expression one text that its parser reads back, for the wire and every listing by @Slonegg in #1340
- Give a statement body an exact text form and store every body and the module's tests as that text by @Slonegg in #1344
- Declare each statement kind with one decorator that writes its protocol, and plan the module's move to Rust by @Slonegg in #1351
- Rename the statement and expression bases to Statement and Expression, and print Kernel IR kinds through their declarations by @Slonegg in #1357
- State every test body as text and let Load and Write take their fields as declared by @Slonegg in #1363
- Attend per key-value head in native's decode and verify steps: one cache read per group, a split combined over its partials by @6erun in #1329
- Price native attention per range of windows, and rank and bench each window's rows at that window by @6erun in #1352
- Copy native attention's FP8 key-value cache into shared memory before the tensor cores read it by @6erun in #1366
- Stage a matmul operand on a vector copy only when every copy starts on its own width by @6erun in #1368
- Pin fast math off in the NVFP4 pair-lookup oracle test, and say -O1 can change a pick by @6erun in #1370
- Merge a computed operand's reads into one 16-byte load only where the whole address is aligned by @6erun in #1369
- Record the profile of Qwen3.8-Flash-Next's first token and eight requests on the V100 by @6erun in #1367
- Fix the nightly refresh: drop the export flag that has killed its prior job and the freeze fold since October 15 by @trifonova-n in #1347
- Drop the freeze rows the compiler no longer realizes each night, with a new emmy golden realize command by @trifonova-n in #1364
- Fold the 454 tune files the nightly never folded into the measurement freeze, and drop the 444 rows the compiler no longer realizes by @trifonova-n in #1356
- Move Loop IR statements and normalization to a Rust crate, with one operation per definition as the normal form by @Slonegg in #1365
- FC-1574 Enumerate the cut pass exact ranges in the Rust crate, not point by point in Python by @Slonegg in #1373
- Name a body's two readings of its states: the states it defines, and the ones that escape it by @Slonegg in #1375
- Measure each recipe's full-context request at onboarding and keep it under the 290 s Relay allows by @trifonova-n in #1323
- Record Qwen3.8-27B-NVFP4's full-context latency on one RTX 5090 and check its output at long context by @trifonova-n in #1382
- Bump Emmy version to 0.3.24 by @trifonova-n in #1383
- Compile Qwen3.8-Flash-Next's fixed-slot expert program again and re-measure its V100 golden by @6erun in #1377
- Run the routed experts once for all eight requests in a Qwen3.8-Flash-Next decode step on the V100 by @6erun in #1378
- Measure the host time of a Qwen3.8-Flash-Next V100 decode step against the fork and propose async scheduling by @6erun in #1384
- One precision knob per math instruction class, and one exponential for both softmax rescale factors by @Slonegg in #1380
- One normal form for indices, carried states and lambdas by @Slonegg in #1379
- Let deploy cloud rent on a named CloudRift node and instance type by @trifonova-n in #1387
- Give an online softmax's rescal...
v0.3.23
What's Changed
- Call Tensor IR to Loop IR conversion lowering to loops by @Slonegg in #1150
- Simplify the README around user examples by @Slonegg in #1151
- Restamp keeps a shared piece and its measurement with the decisions that still mint it, whatever the file order by @gnull in #1138
- Record a route the golden lacks with --bench then --record-greedy, with no pinned decisions or schedules by @gnull in #1152
- Stop renaming a whole loop for a name it rebinds inside, so the outer value's earlier reads keep their name by @gnull in #1153
- Compute an f16/bf16 conv1d at f32 and convert once, like PyTorch by @gnull in #1135
- Release audit compiles each program in the cache regime serving picks for it by @gnull in #1156
- Record the cut decisions of a one-kernel golden bench, so a piece's own cut can be recorded by @gnull in #1157
- Stop offering schedules whose nested tile the binder drops, so one refused site no longer recompiles the whole program by @gnull in #1159
- Refresh golden-bench kernel results and recover missing measurements by @Slonegg in #1155
- Speed up one-token Qwen layers on A100 and H100 by @Slonegg in #1161
- Re-record the Qwen3.8 NVFP4 RTX 5090 golden under main's compiler, and keep the recipe maintained by @gnull in #1160
- Keep each weight in one storage layout across a program and its width twins by @gnull in #1158
- Improve A100 and H100 prefill performance by @Slonegg in #1162
- Optimize A100 causal and grouped-query prefill by @Slonegg in #1163
- Reduce wasted work in A100 causal decode by @Slonegg in #1164
- Reduce split overhead in long A100 decode by @Slonegg in #1165
- Reduce A100 decode overhead and record H100 profiling by @Slonegg in #1166
- Remove the common-kernel benchmark study from main by @Slonegg in #1168
- Plan serving Qwen3.8-Flash-Next on V100 with Emmy's kernels at parity with the 1Cat fork by @6erun in #1167
- Remove host replay gaps from captured benchmark comparisons by @Slonegg in #1169
- Record step 1 of the Qwen3.8-Flash-Next V100 plan: the fork serves it at reference quality by @6erun in #1173
- Return no greedy token when any logit is infinite or NaN by @6erun in #1174
- Bind fixed FP32 scalars and rebuild golden evidence by @Slonegg in #1170
- Fix the compiler workarounds the October audit named by @Slonegg in #1177
- Record tensor-core schedules for the width-1 NVFP4 MLP down projection and GDN out_proj on the RTX 5090 by @gnull in #1172
- Remove node addresses from experiment archives, results and tests by @6erun in #1179
- Record step 2 of the Qwen3.8-Flash-Next V100 plan: the serving shape and the fork baseline by @6erun in #1178
- One Affine statement for every sum, difference, negation and literal product in Loop IR by @Slonegg in #1175
- Ensure GitHub CLI is available for model onboarding by @yash-asvi in #816
- Compile programs for the CPU with LLVM and run them on the Rust runtime by @yash-asvi in #1061
- Run Affine statements on the CPU backend by @Slonegg in #1184
- A cut piece binds only the weights its kernel reads, so a narrower twin's layout settles the prefill twin by @Slonegg in #1181
- Record the GDN and DeepSeek golden routes the Affine normal form left owed by @Slonegg in #1182
- Restamp re-spells a moved cut seam by the pieces it minted by @Slonegg in #1180
- Spell a Loop IR value's coordinates once per body, not once per use by @Slonegg in #1183
- Make the Loop IR normal form a fixed point by @Slonegg in #1186
- Re-measure the Voxtral Mini 3B V100 SXM3 golden after the Affine normal form by @Slonegg in #1188
- Trace every Qwen3.8-Flash-Next layer kind on Transformers 5.19 by @6erun in #1185
- Cut Qwen3.8-Flash-Next layers at the hyper-connection seam so the fork's mixers sit between Emmy's programs by @6erun in #1190
- Plan MoE experts in host memory, and measure OLMoE expert skew on the RTX 5090 by @Slonegg in #1187
- Load Qwen3.8-Flash-Next's NVFP4 routed experts packed into Emmy's expert programs by @6erun in #1192
- Read chosen layers' MoE experts from host memory, and measure the cost on the RTX 5090 by @Slonegg in #1191
- Walk the placement dataset one seam at a time, the way a cold deploy cuts by @trifonova-n in #1189
- Keep the rows a tune measures in the repository freeze and report the prior's regret over them by @trifonova-n in #1176
- Keep each MoE layer's most selected experts on the GPU and the rest in host memory, and measure TPOT against the GPU share by @Slonegg in #1195
- Build the 1Cat V100 image from v1.5.1 and bake its kernel caches so a boot compiles nothing by @6erun in #1196
- A tune writes its rows into the measurement freeze and the nightly refresh folds them by @trifonova-n in #1193
- Tune three Qwen3.8-27B-FP8 kernels on a V100 SXM2 and record the rows by @trifonova-n in #1197
- Use the GPU experts of a MoE layer as a cache of recently routed experts, and measure it with router prediction against the static hot set by @Slonegg in #1198
- Tune fourteen Qwen3.8-27B-FP8 kernels on a V100 SXM2 and record the rows by @trifonova-n in #1200
- Run MoE experts on the CPU near memory speed: a CPU schedule, CPU benching, and NVFP4 on the CPU by @Slonegg in #1199
- Split Qwen3.8-Flash-Next's MoE across tensor-parallel ranks and load its shared expert by @6erun in #1201
- Restamp: compare arms through seam aliases, pair re-keyed children by what they read and write by @gnull in #1194
- Read each GDN1 input-projection weight once per decode step by @gnull in #1171
- Measure running uncached MoE experts on the CPU (strategy G): it loses at one stream on an x16 link by @Slonegg in #1202
- Fail a serving boot that meets a program which never finishes, naming the program by @6erun in #1204
- Host the 1Cat fork's sparse attention inside Emmy's Qwen3.8-Flash-Next model by @6erun in #1203
- Serve MoE experts through generic paging: compiled paging programs over paged stacked weights by @Slonegg in #1205
- Serve GDN layers without reading vLLM metadata back from the GPU each step by @gnull in #1209
- Capture CUDA graphs of decode steps on GDN models by @gnull in #1211
- Price a weight-layout decision for a reader still to cut by the cut recorded on it by @6erun in #1207
- Form every cut piece from its own body, so one identity is one kernel by @6erun in #1206
- Run prefill steps just above the decode bucket on the chunk twin when the chunk is narrow by @gnull in #1214
- Let GDN programs write their new state straight into vLLM's cache block by @gnull in #1215
- Tile the weight column a direct read can reach when a matmul's weight owns several axes by @6erun in #1208
- Reproduce the V100 SXM2 golden's GEMV split by measuring it in the Llama golden and refitting the placement prior by @6erun in #1217
- Restamp a moved seam whose cut also re-keys a piece instead of dropping it by @6erun in #1222
- Lay out a matmul piece's workspace in its weight's order by @6erun in #1221
- Order a kernel's grid by its store's strides, the same in its program and alone by @6erun in #1216
- Serve Qwen3.8-27B-NVFP4 on the RTX 5090 from measured kernel choices only, and let the strict correctness check allow one rounding step by @g...
v0.3.22
What's Changed
- Ship goldens only with bundled recipes and learn weight layouts in the placement prior by @Slonegg in #1148
- Bump Emmy version to 0.3.22 by @Slonegg in #1149
Full Changelog: v0.3.21...v0.3.22
v0.3.21
What's Changed
- Fix the cut-pass crash on fused sibling producers (gdn16) by @gnull in #1107
- Stage a unit N side for the native FP4 tile: one-token NVFP4 linears take the tensor core by @gnull in #1108
- Restore staged child pin tests after projection sharing by @gnull in #1109
- Define the FP4 encode helper for fragment epilogues and read each element's block scale by @gnull in #1111
- Refuse a serial lowering that discards a selected nested tile, and correct the test records that relied on it by @gnull in #1112
- Declare a split or layout arm's output redirects in the parent, so parallel fork expansion keeps every output by @gnull in #1114
- Keep the placement prior on the search's last resolve, so an exhausted schedule search does not fall back to one fused kernel by @gnull in #1116
- Refuse to record any flagged bench row, in both record modes by @gnull in #1117
- Normalize negative slice starts in the Torch reference evaluator, and surface reference failures by @gnull in #1118
- Enable MMA tiles for rank-one reduction projections by @gnull in #1100
- Refuse a cross-CTA split that would narrow a contraction's whole-row statistic by @gnull in #1124
- Stage a symbolic token row for the native FP4 tile: any-width NVFP4 linears take the tensor core by @gnull in #1110
- Draw the synthesized NVFP4 scales in proportion to the synthesized weights and activations by @gnull in #1123
- Type a cut workspace from its producer's value, so a later FP8 output encode cannot narrow a BF16 operand by @gnull in #1115
- Read each GDN layer's KV-cache block indices once per model forward, not once per layer by @gnull in #1127
- Restamp a golden's routing entries parents first, whatever order the file stores them in by @gnull in #1129
- Re-record and tune the Qwen3.8 NVFP4 RTX 5090 golden: strict coverage of all nine serving programs, 26 tokens per second served by @gnull in #1120
- Train the placement prior on the complete recorded cut, not on its subsets, when the cut pass offers it by @gnull in #1119
- Add an opt-in cold-cache bench mode: evict L2 before every timed launch, keep cold rows in their own regime by @gnull in #1128
- Refuse a cut seam whose producer only loads, so the cut-workspace ordering sees no self-cycle by @gnull in #1130
- Let a priced layout decision price the fuse arm at its placement fork, so a recorded layout route replays under strict evidence by @gnull in #1131
- Make the strict whole-program check adopt the candidate's quantized values at the quantize nodes by @gnull in #1132
- Place a unit row for rowless contractions whose columns span several free axes, so the GDN input projections take the MMA tile by @gnull in #1133
- Record how DeepSeek V4 on V100 loses details in long prompts, alike on the Emmy image and the plain fork by @6erun in #1126
- Select each serving program’s recorded cache regime by @gnull in #1136
- Derive kernel identity and the Loop IR normal form from value numbering by @Slonegg in #1106
- Replay nested golden cuts on fresh children and keep composed cuts open by @Slonegg in #1140
- Keep accumulator scope, bound coordinates and free-axis roles in the normal form by @Slonegg in #1141
- Check copied-operand cover when the compute fill builds its slabs by @Slonegg in #1142
- Regenerate the sm120 serving fixture under the current normal form by @Slonegg in #1143
- Carry the Qwen prefill and decode recordings over from the replay recovery by @Slonegg in #1144
torch.compile()backend, functionalize writes to inputs, and fix bugs by @apaz-cli in #1145- Restore fast 3584-wide linear picks on RTX 5090 by @Slonegg in #1146
- Bump Emmy version to 0.3.21 by @Slonegg in #1147
New Contributors
Full Changelog: v0.3.20...v0.3.21
v0.3.20
What's Changed
- Give the image pull two hours, because a 164 GB serving image needs about one by @6erun in #1121
- Bump version from 0.3.19 to 0.3.20 by @6erun in #1122
Full Changelog: v0.3.19...v0.3.20
v0.3.19
What's Changed
- Restart an engine container that stays up after its engine dies by @Slonegg in #1071
- Serve FP8 checkpoints with dynamic activations from their coded weights by @Slonegg in #1085
- Serve Voxtral Mini 3B with audio input on one V100 by @Slonegg in #1095
- Native export keeps an FP8 checkpoint's trunk coded by @Slonegg in #1094
- Enable tensor-core schedules for batched matrix-vector projections by @gnull in #1086
- Preserve cut-child address across projection split by @gnull in #1087
- Reuse NVFP4 output projection across encode layouts by @gnull in #1089
- Preserve integer values in matrix epilogues by @gnull in #1101
- Rank a cut's arms without pricing unmeasured pieces, and build them on forked workers by @Slonegg in #1103
- Install python3.12-venv before setup in the nightly golden fill by @Slonegg in #1104
- Re-measure and re-tune the hardware goldens on five GPUs, and refit the priors by @Slonegg in #1098
- Replay later placement cuts on the same child by @gnull in #1099
- Reuse gate and up projections across Qwen block and pair consumers by @gnull in #1102
- Fix the placement prior on V100 K-splits, and keep the prior reproduction gate green by @Slonegg in #1105
- Bump Emmy package version to 0.3.19 by @Slonegg in #1113
Full Changelog: v0.3.18...v0.3.19
v0.3.18
What's Changed
- Roll an unrolled chain only when fusing it would recompute a state by @Slonegg in #1065
- Benchmark a shared prompt prefix, and measure the prefix cache on the DeepSeek V4 Emmy image by @6erun in #1068
- Serve mixed base and LoRA Llama requests through Emmy by @Slonegg in #1064
- Record faster Qwen3.8 NVFP4 solve and encoding goldens on RTX 5090 by @gnull in #1069
- Reduce V100 FP16 prefill layer time with measured gate/up schedules by @Slonegg in #1063
- List compiler gaps and fill hardware goldens nightly by @Slonegg in #1075
- Record FP32 hardware goldens so the greedy picks fast FP32 schedules by @Slonegg in #1073
- Pin Qwen3.8-27B-FP8's V100 key/value cache at 4.5 GiB so a restart cannot shrink its memory margin by @Slonegg in #1072
- Read only hardware and maintained-recipe goldens, and let a person lock a recipe's lifecycle by @Slonegg in #1079
- Form cut pieces as the kernel their own program forms by @Slonegg in #1077
- Autotune a kernel with emmy run --tune, one kernel at a time with --kernel by @Slonegg in #1076
- Prefix-cache results: a cache hit computes exactly what a split prefill computes by @6erun in #1070
- Re-record the golden routes #1077 re-keyed or dropped by @Slonegg in #1081
- Allow explicit staged placement cuts on remaining consumers by @gnull in #1074
- Record a fast RTX 5090 GDN64 input route for Qwen3.8 NVFP4 by @gnull in #1082
- Nightly: fix the failures of the 2026-10-06 and 2026-10-07 runs by @Slonegg in #1084
- DeepSeek V4 on V100: the recipe serves the Emmy image at the full 1M context by @6erun in #1088
- Keep one serving-lane golden per compute capability, and add the V100 one by @Slonegg in #1090
- V100 LoRA serving at stock decode speed: split-K decode schedules and an FP64 strict check by @Slonegg in #1080
- Beat torch.compile on V100 FP16 decode with an exact one-key softmax fold and vector weight reads by @Slonegg in #1078
- Fix the bot's push-to-main helper failing on macOS's bash 3.2 by @6erun in #1091
- Profile the V100 LoRA 512-row post-attention program and re-record its golden rows by @Slonegg in #1093
- Refit the schedule prior on the FP32 hardware goldens by @Slonegg in #1092
- Trace every part of Qwen3-Omni: decoder-stack selection, gated shared expert, N-way cat, transposed and 2-D convolutions by @Slonegg in #1096
- Bump version from 0.3.17 to 0.3.18 by @6erun in #1097
Full Changelog: v0.3.17...v0.3.18
v0.3.17
What's Changed
- Share K/V producers to reduce Qwen decode latency by @Slonegg in #1011
- Serve attention layers that carry an output gate (Qwen3.5 full attention) by @gnull in #1022
- Bind the folded 1 + weight norm constant in serving plans by @gnull in #1024
- Serve dense generative trunks in BF16 by @gnull in #1025
- Refresh model lifecycle and onboarding by @deplodock-runner-bot[bot] in #1021
- Recognize contractions beside unrelated reduction states by @Slonegg in #1028
- Optimize Qwen3.8 V100 kernels with measured manual schedules by @Slonegg in #1029
- Share prefill K/V kernels and normalize V100 projection inputs by @Slonegg in #1019
- DeepSeek V4 on V100: capture decode batches up to the decode bucket by @6erun in #1020
- Fold model discovery into the nightly refresh and share its git and Discord plumbing by @Slonegg in #1026
- Reduce V100 decode latency with an existing one-warp schedule by @Slonegg in #1031
- Export golden pools in parallel, one pool per worker process by @Slonegg in #1033
- Document H100 optimization results and rejected candidates by @Slonegg in #1034
- Keep Qwen3.8 triangular updates in shared memory on V100 by @Slonegg in #1032
- Expand placement training with Qwen3 hardware goldens by @Slonegg in #1038
- Run every test session on a fresh tune DB and pin the SDPA tensor-core test to its kernel by @6erun in #1035
- DeepSeek V4 on 16 V100s: rerun the A/B with the Emmy image rebuilt from current main by @6erun in #1036
- Refit schedule prior and extend CI limit for hardware goldens by @Slonegg in #1041
- Record a Qwen3.8-27B-NVFP4 golden on an RTX 5090 by @gnull in #1027
- One featurizer for both priors, and a placement gate walk that stops at the kernel's own fork by @Slonegg in #1045
- Run a routed MoE layer's experts in one runtime call: DeepSeek V4 on 16 V100s reaches its first token before the 1Cat fork by @6erun in #1039
- Trace static-FP8 checkpoints, serve them coded, and add the Ministral 3 14B golden for the RTX 5090 by @Slonegg in #1044
- Refuse a cross-CTA split that takes a whole-row statistic over one slice by @Slonegg in #1046
- Add pinned LoRA serving and qualify Llama 3.1 on V100 by @Slonegg in #1043
- Serve the GDN layers of Qwen3.5 / Qwen3.8 hybrid models through emmy by @gnull in #1023
- Compute a kernel's identity and features where they are read; store neither on ops, in goldens or in the tune DB by @Slonegg in #1047
- DeepSeek V4 on V100: cooperative reduces for the 4,096-token pre program (2.9 to 0.8 ms per layer) by @6erun in #1049
- Stabilize nightly discovery and validate placement export by @Slonegg in #1050
- Fuse a reduction that reads at an index naming a symbolic dim, so the DeepSeek V4 symbolic post routing target compiles standalone by @6erun in #1048
- Serve Qwen3.5 text models natively: carried Gated DeltaNet state across steps by @Slonegg in #1042
- Restore the pull-request template and fail the suite when a PR body is written over it by @6erun in #1051
- Choose transposed weight layout from measured kernel evidence by @Slonegg in #1052
- DeepSeek V4 on 16 V100s: A/B rerun at main 992de5c, Emmy faster on one request and equal at 8 concurrent by @6erun in #1056
- Answer transport-independent schedule checks once per choice and build each cut arm once by @Slonegg in #1054
- DeepSeek V4 on V100: record the published Emmy serving image and add its single-variant recipe by @6erun in #1057
- Retire the DeepSeek V4 V100 serving plan and keep its lessons in the durable notes by @6erun in #1058
- Improve V100 decode schedule and refresh FP16 layer evidence by @Slonegg in #1060
- Record and retune Llama 3.1 8B compiler golden on V100 by @Slonegg in #1055
- Decide every kernel-set fork before any schedule, from what its arms are by @Slonegg in #1059
- Retune Llama 3.1 post-attention M1 on V100 by @Slonegg in #1062
- Declare image input per recipe and export it in the catalog, with an image smoke test by @trifonova-n in #1053
- Draw large schedule pools seven times faster and sample 8192 rows by default by @Slonegg in #1066
- Declare image input per deployment, and keep the vision tower on Qwen3.8-27B-FP8's H200 share by @trifonova-n in #1067
Full Changelog: v0.3.16...v0.3.17