Skip to content

Enable ONNX external-data sharding - #503

Open
justinchuby wants to merge 14 commits into
mainfrom
parallel-external-data-save
Open

Enable ONNX external-data sharding#503
justinchuby wants to merge 14 commits into
mainfrom
parallel-external-data-save

Conversation

@justinchuby

@justinchuby justinchuby commented Aug 16, 2026 •
edited
Loading

Copy link
Copy Markdown
Member

Summary

  • Raise the minimum supported onnx_ir version from 0.2.1 to 1.0.0.
  • Forward --max-shard-size to ONNX external-data saves as max_shard_size_bytes, an API already available in onnx_ir 1.0.0.
  • Make the save progress callback thread-safe. With newer onnx_ir versions it renders one fixed-position tqdm bar per shard; with onnx_ir 1.0.0 it falls back to one global bar.
  • Keep Mobius independent of the pending parallel-save API: Mobius does not expose or pass max_workers.

Validation

The ONNX sharding round-trip and callback tests pass against both the published onnx_ir==1.0.0 package and the current onnx/ir-py#476 branch.

@justinchuby
Save external data with multiple threads
e0755fb 
Weights that need a dtype cast are held as ir.LazyTensor and converted at
serialization time, so saving alternates between casting a tensor and
writing it, leaving both the CPU and the disk idle half the time. onnx_ir
gained a max_workers option that overlaps materialization with disk writes
and parallelizes both, with peak memory bounded independently of the worker
count.

Default ModelPackage.save to 8 workers and thread the option through. The
option is feature-detected because it is not in onnx_ir 0.2.x, which the
package still supports, so older installs keep saving serially instead of
raising TypeError.

Fix the progress bar for concurrent saving. onnx_ir serializes callbacks
with a lock but no longer delivers them in index order, and the bar mutated
total/set_description without any synchronization of its own. Guard the
closure state with a lock and keep counting invocations rather than
tracking indices.

Measured on Qwen2.5-7B cast from bfloat16 to float16, 15.26GB of external
data: 264.2s serial -> 176.0s with 8 workers (1.50x). 4 workers gives 1.45x
and 16 gives 1.48x, so the default of 8 is near the knee.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40

Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
justinchuby requested review from a team and a lite review from Copilot August 16, 2026 01:45
@justinchuby
Require onnx_ir>=1.1.0 instead of feature-detecting max_workers
1383495 
max_workers ships in onnx_ir 1.1.0, so depend on it directly and drop the
inspect.signature fallback. Merge this after 1.1.0 is released.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40

Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
Derive the default save worker count instead of hardcoding 8
7309783 
Hardcoding 8 happened to match this machine's performance-core count, which
was a coincidence rather than a reason.

Measuring the phases separately shows the bottleneck is the storage device,
not the CPU: casting bf16 to f16 runs at 12.2 GB/s while serial writes reach
only 1.6 GB/s. A write-only scaling test saturates at 2 threads (2.1x) and
does not improve through 24. So the speedup comes from overlapping the cast
with the write, not from write parallelism.

That is why the default deliberately does not scale with the core count --
something like cpu_count/2 would oversubscribe against torch's own intra-op
pool (already 8 threads here) without making the disk faster. Keep a small
constant, but clamp it by os.cpu_count() so small containers do not spawn
more threads than they can run, and handle cpu_count() returning None.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40

Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
Copilot AI reviewed Aug 16, 2026
View reviewed changes

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR improves ModelPackage.save() performance when saving ONNX external data by enabling concurrent materialization + disk writes via onnx_ir’s new max_workers option when available, while preserving compatibility with older onnx_ir versions through feature detection. It also updates the progress callback behavior and adds test coverage for serial/parallel equivalence and callback ordering/thread-safety.

Changes:

  • Add max_workers (default 8) to ModelPackage.save() and pass it to ir.save(..., max_workers=...) when supported.
  • Add runtime feature detection for ir.save(max_workers=...) to avoid breaking older onnx_ir installs.
  • Add tests validating byte-identical output, load roundtrip, callback correctness under out-of-order and concurrent invocations, and feature detection behavior.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File Description
src/mobius/_model_package.py Adds max_workers support with feature detection; makes progress callback state thread-safe.
src/mobius/_model_package_test.py Adds TestParallelSave verifying serial/parallel equivalence and progress-callback correctness.
Suppressed comments (1)

src/mobius/_model_package_test.py:163

  • This test relies on a specific __closure__ cell index to fetch the tqdm bar. Closure ordering is an implementation detail and can change with refactors/Python versions, making the test fragile. Extract the bar by searching closure cells for the object that has tqdm methods (e.g., set_description) instead of indexing.
                _Tensor(),

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Comment thread src/mobius/_model_package_test.py Outdated
),
)
# Closure state is private to the callback; assert via the bound bar.
bar = callback.__closure__[1].cell_contents
Comment thread src/mobius/_model_package.py
Comment on lines +308 to +312
with lock:
if not total_set:
pbar.total = metadata.total
total_set = True
pbar.update()
justinchuby and others added 3 commits August 15, 2026 19:05
@justinchuby
Do not shrink the default worker count on small machines
8e57e17 
Sizing the pool by core count gets the tradeoff backwards. With torch pinned
to one intra-op thread, i.e. a small or slow machine, 8 workers is the
fastest configuration (1.55x) while 2 workers is slower than serial: more
concurrency is needed to hide a slow serial cast, not less. cpu_count() // 2
would give 1 on a 2-core box, disabling the optimization exactly where it
helps most.

Oversubscription is not penalized in practice either. With full torch
threads, everything from 4 to 32 workers lands within noise (1.37x-1.45x),
because these threads block in tofile with the GIL released rather than
competing for CPU.

Drop the core-count clamp and keep a plain constant.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40

Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
Note that accelerator-resident weights benefit most
7d3fafb 
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40

Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
Enable sharding for ONNX external data
91996c9 
Forward max_shard_size_bytes to ir.save instead of limiting the option to
safetensors. Update the CLI help and add a behavioral test that writes,
discovers, and reloads three ONNX external-data shards.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
justinchuby force-pushed the parallel-external-data-save branch from 70350a0 to 91996c9 Compare August 16, 2026 04:29
@justinchuby
Keep Mobius compatible with onnx_ir 1.0
09e664b 
Do not expose or pass max_workers or ONNX max_shard_size_bytes from
Mobius. Those APIs belong to the pending onnx_ir change, and calling them
would make Mobius fail before that change is released. Require only
onnx_ir 1.0.0 so the Mobius update is independent of the ir-py PR.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby justinchuby changed the title Save external data with multiple threads Require onnx_ir 1.0.0 Aug 16, 2026
@justinchuby
Enable ONNX sharding with onnx_ir 1.0
1348af4 
Forward max_shard_size_bytes for ONNX external data, an API already
available in onnx_ir 1.0.0. Keep Mobius independent of the pending
parallel-save change by not exposing or passing max_workers.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby justinchuby changed the title Require onnx_ir 1.0.0 Enable ONNX external-data sharding Aug 16, 2026
@github-actions

github-actions Bot commented Aug 16, 2026 •
edited
Loading

Copy link
Copy Markdown

Performance Comparison

Comparing 55dd674 → e4ae9e9

Model Metric Baseline Current Delta
bert (feature-extraction) model_size_bytes 359 KB 359 KB +0.0% ⚪
bert (feature-extraction) num_nodes 60 60 +0.0% ⚪
falcon model_size_bytes 364 KB 364 KB +0.0% ⚪
falcon num_nodes 66 68 +3.0% ⚪
gemma2 model_size_bytes 428 KB 428 KB +0.0% ⚪
gemma2 num_nodes 105 107 +1.9% ⚪
gpt2 model_size_bytes 388 KB 388 KB +0.0% ⚪
gpt2 num_nodes 54 54 +0.0% ⚪
llama model_size_bytes 425 KB 425 KB +0.0% ⚪
llama num_nodes 60 62 +3.3% ⚪
llama (static-cache) model_size_bytes 425 KB 425 KB +0.0% ⚪
llama (static-cache) num_nodes 56 58 +3.6% ⚪
mamba (ssm-text-generation) model_size_bytes 296 KB 296 KB +0.0% ⚪
mamba (ssm-text-generation) num_nodes 94 98 +4.3% ⚪
phi3 model_size_bytes 421 KB 421 KB +0.0% ⚪
phi3 num_nodes 58 60 +3.4% ⚪
phi3 (static-cache) model_size_bytes 421 KB 421 KB +0.0% ⚪
phi3 (static-cache) num_nodes 54 56 +3.7% ⚪
qwen2 model_size_bytes 425 KB 425 KB +0.0% ⚪
qwen2 num_nodes 60 62 +3.3% ⚪
qwen2 (static-cache) model_size_bytes 425 KB 425 KB +0.0% ⚪
qwen2 (static-cache) num_nodes 56 58 +3.6% ⚪
qwen3_5_moe (hybrid-text-generation) model_size_bytes 506 KB 506 KB +0.0% ⚪
qwen3_5_moe (hybrid-text-generation) num_nodes 264 275 +4.2% ⚪
qwen3_5_text (hybrid-text-generation) model_size_bytes 458 KB 458 KB +0.0% ⚪
qwen3_5_text (hybrid-text-generation) num_nodes 126 129 +2.4% ⚪
qwen3_5_vl (hybrid-qwen-vl) model_size_bytes 977 KB 977 KB +0.0% ⚪
qwen3_5_vl (hybrid-qwen-vl) num_nodes 428 431 +0.7% ⚪
t5 (seq2seq) model_size_bytes 836 KB 836 KB +0.0% ⚪
t5 (seq2seq) num_nodes 166 166 +0.0% ⚪
whisper (speech-to-text) model_size_bytes 1008 KB 1008 KB +0.0% ⚪
whisper (speech-to-text) num_nodes 128 128 +0.0% ⚪

No performance regressions.

@github-actions

github-actions Bot commented Aug 16, 2026 •
edited
Loading

Copy link
Copy Markdown

🏗️ Architecture Diff

Comparing 55dd674 → e4ae9e9

Model Sub-model Changes Status
bert (feature-extraction) model 0 ⚪
falcon model 8 🟡
gemma2 model 12 🟡
gemma4 (gemma4) decoder 55 🟡
gemma4 (gemma4) embedding 2 🟡
gemma4 (gemma4) vision_encoder 0 ⚪
gemma4_text model 55 🟡
gpt2 model 0 ⚪
llama model 8 🟡
llama (static-cache) model 9 🟡
mamba (ssm-text-generation) model 23 🟡
phi3 model 9 🟡
phi3 (static-cache) model 10 🟡
qwen model 8 🟡
qwen (static-cache) model 9 🟡
qwen2 model 8 🟡
qwen2 (static-cache) model 9 🟡
qwen2_moe model 295 🟡
qwen2_moe (static-cache) model 295 🟡
qwen3 model 9 🟡
qwen3 (static-cache) model 10 🟡
qwen3_5_moe (hybrid-text-generation) model 299 🟡
qwen3_5_text (hybrid-text-generation) model 18 🟡
qwen3_5_vl (hybrid-qwen-vl) decoder 19 🟡
qwen3_5_vl (hybrid-qwen-vl) embedding 0 ⚪
qwen3_5_vl (hybrid-qwen-vl) vision_encoder 0 ⚪
qwen3_moe model 236 🟡
qwen3_moe (static-cache) model 40 🟡
qwen3_next (hybrid-text-generation) model 615 🟡
t5 (seq2seq) decoder 0 ⚪
t5 (seq2seq) encoder 0 ⚪
whisper (speech-to-text) decoder 0 ⚪
whisper (speech-to-text) encoder 2 🔵
falcon / model — 8 change(s)

Op summary: 66 → 68 nodes

--- base
+++ head
@@ -34,7 +34,8 @@
 LayerNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Add
@@ -57,7 +58,8 @@
 LayerNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Add

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[51] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[51] RotaryEmbedding: input_ids [92, 45, 46] → [90, 45, 46]
gemma2 / model — 12 change(s)

Op summary: 105 → 107 nodes

--- base
+++ head
@@ -54,7 +54,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -87,7 +88,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[77] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[65] Add: input_ids [32, 17] → [103, 116]
  • node[77] RotaryEmbedding: input_ids [124, 52, 53] → [122, 52, 53]
  • node[82] Add: input_ids [41, 17] → [135, 40]
  • node[85] Add: input_ids [42, 17] → [117, 138]
  • node[97] Add: input_ids [138, 150] → [46, 17]
gemma4 (gemma4) / decoder — 55 change(s)

Op summary: 125 → 127 nodes

--- base
+++ head
@@ -64,7 +64,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -107,7 +108,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -119,7 +121,7 @@
 RMSNormalization
 Transpose
 MatMul
-Constant
+CastLike
 Div
 Tanh
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + CastLike

Removed nodes:

  • - Swish
  • - Swish
  • - Constant

Connectivity changes:

  • node[5] Unsqueeze: input_ids [53, 12] → [54, 12]
  • node[8] Sub: input_ids [56, 55] → [57, 56]
  • node[9] Slice: input_ids [53, 57, 56, 12] → [54, 58, 57, 12]
  • node[10] Unsqueeze: input_ids [58, 13] → [59, 13]
  • node[11] GreaterOrEqual: input_ids [59, 54] → [60, 55]
  • node[13] Cast: input_ids [61] → [62]
  • node[14] And: input_ids [62, 60] → [63, 61]
  • node[15] Where: input_ids [63, 14, 15] → [64, 14, 15]
  • node[16] Cast: input_ids [64] → [65]
  • node[17] Unsqueeze: input_ids [65, 12] → [66, 12]
  • node[19] Unsqueeze: input_ids [67, 12] → [68, 12]
  • node[22] Sub: input_ids [70, 69] → [71, 70]
  • node[23] Slice: input_ids [67, 71, 70, 12] → [68, 72, 71, 12]
  • node[24] Unsqueeze: input_ids [72, 13] → [73, 13]
  • node[25] GreaterOrEqual: input_ids [73, 68] → [74, 69]
  • node[27] Cast: input_ids [75] → [76]
  • node[28] And: input_ids [76, 74] → [77, 75]
  • node[29] Where: input_ids [77, 14, 15] → [78, 14, 15]
  • node[30] Cast: input_ids [78] → [79]
  • node[31] Unsqueeze: input_ids [79, 12] → [80, 12]
  • node[34] MatMul: input_ids [81, 82] → [82, 83]
  • node[35] Reshape: input_ids [83, 19] → [84, 19]
  • node[36] RMSNormalization: input_ids [84, 20] → [85, 20]
  • node[37] Reshape: input_ids [85, 21] → [86, 21]
  • node[38] RotaryEmbedding: input_ids [86, 49, 50] → [87, 50, 51]
  • node[40] MatMul: input_ids [81, 88] → [82, 89]
  • node[41] Reshape: input_ids [89, 19] → [90, 19]
  • node[42] RMSNormalization: input_ids [90, 23] → [91, 23]
  • node[43] Reshape: input_ids [91, 21] → [92, 21]
  • node[44] RotaryEmbedding: input_ids [92, 49, 50] → [93, 50, 51]
  • node[46] MatMul: input_ids [81, 94] → [82, 95]
  • node[48] Reshape: input_ids [95, 96] → [96, 97]
  • node[49] Cast: input_ids [97] → [98]
  • node[50] Mul: input_ids [98, 98] → [99, 99]
  • node[51] ReduceMean: input_ids [99, 25] → [100, 25]
  • node[53] Add: input_ids [100, 101] → [101, 102]
  • node[54] Sqrt: input_ids [102] → [103]
  • node[55] Div: input_ids [98, 103] → [99, 104]
  • node[56] CastLike: input_ids [104, 97] → [105, 98]
  • node[57] Reshape: input_ids [105, 21] → [106, 21]
  • node[58] Attention: input_ids [87, 93, 106, 66, 3, 4] → [88, 94, 107, 67, 3, 4]
  • node[60] MatMul: input_ids [107, 110] → [108, 111]
  • node[61] RMSNormalization: input_ids [111, 27] → [112, 27]
  • node[62] Add: input_ids [0, 112] → [0, 113]
  • node[63] RMSNormalization: input_ids [113, 28] → [114, 28]
  • node[65] MatMul: input_ids [114, 115] → [115, 116]

Initializer changes:

  • count 42 → 43; dtype distribution: FLOAT: 36 → 37
gemma4 (gemma4) / embedding — 2 change(s)

Op summary: 19 → 18 nodes

--- base
+++ head
@@ -1,5 +1,4 @@
 Gather
-Constant
 Mul
 Constant
 Equal

Removed nodes:

  • - Constant

Initializer changes:

  • count 3 → 4; dtype distribution: FLOAT: 1 → 2
gemma4_text / model — 55 change(s)

Op summary: 128 → 129 nodes

--- base
+++ head
@@ -1,5 +1,4 @@
 Gather
-Constant
 Mul
 Gather
 Gather
@@ -67,7 +66,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -110,7 +110,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -122,7 +123,7 @@
 RMSNormalization
 Transpose
 MatMul
-Constant
+CastLike
 Div
 Tanh
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + CastLike

Removed nodes:

  • - Constant
  • - Swish
  • - Swish
  • - Constant

Connectivity changes:

  • node[3] Gather: input_ids [8, 2] → [10, 2]
  • node[4] Gather: input_ids [9, 2] → [11, 2]
  • node[5] Gather: input_ids [10, 2] → [12, 2]
  • node[9] Shape: input_ids [0] → [1]
  • node[23] Shape: input_ids [0] → [1]
  • node[70] Transpose: input_ids [31] → [32]
  • node[71] MatMul: input_ids [118, 122] → [119, 124]
  • node[72] Mul: input_ids [121, 123] → [123, 125]
  • node[73] Transpose: input_ids [32] → [33]
  • node[74] MatMul: input_ids [124, 125] → [126, 127]
  • node[75] RMSNormalization: input_ids [126, 33] → [128, 34]
  • node[76] Add: input_ids [117, 127] → [118, 129]
  • node[77] Mul: input_ids [128, 17] → [130, 18]
  • node[78] RMSNormalization: input_ids [129, 35] → [131, 36]
  • node[79] Transpose: input_ids [36] → [37]
  • node[80] MatMul: input_ids [130, 131] → [132, 133]
  • node[81] Reshape: input_ids [132, 20] → [134, 21]
  • node[82] RMSNormalization: input_ids [133, 37] → [135, 38]
  • node[83] Reshape: input_ids [134, 22] → [136, 23]
  • node[84] RotaryEmbedding: input_ids [135, 55, 56] → [137, 56, 57]
  • node[85] Transpose: input_ids [38] → [39]
  • node[86] MatMul: input_ids [130, 137] → [132, 139]
  • node[87] Reshape: input_ids [138, 20] → [140, 21]
  • node[88] RMSNormalization: input_ids [139, 39] → [141, 40]
  • node[89] Reshape: input_ids [140, 22] → [142, 23]
  • node[90] RotaryEmbedding: input_ids [141, 55, 56] → [143, 56, 57]
  • node[91] Transpose: input_ids [40] → [41]
  • node[92] MatMul: input_ids [130, 143] → [132, 145]
  • node[94] Reshape: input_ids [144, 145] → [146, 147]
  • node[95] Cast: input_ids [146] → [148]
  • node[96] Mul: input_ids [147, 147] → [149, 149]
  • node[97] ReduceMean: input_ids [148, 26] → [150, 27]
  • node[99] Add: input_ids [149, 150] → [151, 152]
  • node[100] Sqrt: input_ids [151] → [153]
  • node[101] Div: input_ids [147, 152] → [149, 154]
  • node[102] CastLike: input_ids [153, 146] → [155, 148]
  • node[103] Reshape: input_ids [154, 22] → [156, 23]
  • node[104] Attention: input_ids [136, 142, 155, 84, 5, 6] → [138, 144, 157, 85, 5, 6]
  • node[105] Transpose: input_ids [41] → [42]
  • node[106] MatMul: input_ids [156, 159] → [158, 161]
  • node[107] RMSNormalization: input_ids [160, 42] → [162, 43]
  • node[108] Add: input_ids [129, 161] → [131, 163]
  • node[109] RMSNormalization: input_ids [162, 43] → [164, 44]
  • node[110] Transpose: input_ids [44] → [45]
  • node[111] MatMul: input_ids [163, 164] → [165, 166]

Initializer changes:

  • count 43 → 45; dtype distribution: FLOAT: 37 → 39
llama / model — 8 change(s)

Op summary: 60 → 62 nodes

--- base
+++ head
@@ -25,7 +25,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -48,7 +49,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[42] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[42] RotaryEmbedding: input_ids [71, 32, 33] → [70, 32, 33]
llama (static-cache) / model — 9 change(s)

Op summary: 56 → 58 nodes

--- base
+++ head
@@ -19,7 +19,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -44,7 +45,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[36] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[36] RotaryEmbedding: input_ids [65, 32, 33] → [64, 32, 33]
  • node[38] TensorScatter: input_ids [5, 67, 6] → [4, 70, 6]
mamba (ssm-text-generation) / model — 23 change(s)

Op summary: 94 → 98 nodes

--- base
+++ head
@@ -9,7 +9,8 @@
 Conv
 Unsqueeze
 Add
-Swish
+Sigmoid
+Mul
 Transpose
 Transpose
 MatMul
@@ -40,7 +41,8 @@
 Mul
 Add
 Unsqueeze
-Swish
+Sigmoid
+Mul
 Mul
 Transpose
 MatMul
@@ -54,7 +56,8 @@
 Conv
 Unsqueeze
 Add
-Swish
+Sigmoid
+Mul
 Transpose
 Transpose
 MatMul
@@ -85,7 +88,8 @@
 Mul
 Add
 Unsqueeze
-Swish
+Sigmoid
+Mul
 Mul
 Transpose
 MatMul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish
  • - Swish
  • - Swish

Modified attributes:

  • node[13] Transpose: perm: [1, 0] → [0, 2, 1]

Connectivity changes:

  • node[13] Transpose: input_ids [17] → [48]
  • node[23] Unsqueeze: input_ids [59, 22] → [58, 21]
  • node[31] Squeeze: input_ids [67, 9] → [64, 9]
  • node[61] Transpose: input_ids [31] → [30]
  • node[62] MatMul: input_ids [102, 105] → [102, 103]
  • node[71] Unsqueeze: input_ids [103, 11] → [113, 22]
  • node[72] Mul: input_ids [111, 115] → [114, 115]
  • node[77] Mul: input_ids [119, 4] → [119, 120]
  • node[79] Squeeze: input_ids [104, 9] → [121, 9]
  • node[84] Mul: input_ids [29, 127] → [125, 127]
phi3 / model — 9 change(s)

Op summary: 58 → 60 nodes

--- base
+++ head
@@ -26,7 +26,8 @@
 Transpose
 MatMul
 Split
-Swish
+Sigmoid
+Mul
 Mul
 Transpose
 MatMul
@@ -48,7 +49,8 @@
 Transpose
 MatMul
 Split
-Swish
+Sigmoid
+Mul
 Mul
 Transpose
 MatMul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[41] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[29] Mul: input_ids [60, 59] → [58, 60]
  • node[41] RotaryEmbedding: input_ids [69, 30, 31] → [68, 30, 31]
phi3 (static-cache) / model — 10 change(s)

Op summary: 54 → 56 nodes

--- base
+++ head
@@ -20,7 +20,8 @@
 Transpose
 MatMul
 Split
-Swish
+Sigmoid
+Mul
 Mul
 Transpose
 MatMul
@@ -44,7 +45,8 @@
 Transpose
 MatMul
 Split
-Swish
+Sigmoid
+Mul
 Mul
 Transpose
 MatMul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[35] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[23] Mul: input_ids [54, 53] → [52, 54]
  • node[35] RotaryEmbedding: input_ids [63, 30, 31] → [62, 30, 31]
  • node[37] TensorScatter: input_ids [5, 65, 6] → [4, 68, 6]
qwen / model — 8 change(s)

Op summary: 60 → 62 nodes

--- base
+++ head
@@ -25,7 +25,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -48,7 +49,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[42] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[42] RotaryEmbedding: input_ids [71, 32, 33] → [70, 32, 33]
qwen (static-cache) / model — 9 change(s)

Op summary: 56 → 58 nodes

--- base
+++ head
@@ -19,7 +19,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -44,7 +45,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[36] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[36] RotaryEmbedding: input_ids [65, 32, 33] → [64, 32, 33]
  • node[38] TensorScatter: input_ids [5, 67, 6] → [4, 70, 6]
qwen2 / model — 8 change(s)

Op summary: 60 → 62 nodes

--- base
+++ head
@@ -25,7 +25,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -48,7 +49,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[42] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[42] RotaryEmbedding: input_ids [71, 32, 33] → [70, 32, 33]
qwen2 (static-cache) / model — 9 change(s)

Op summary: 56 → 58 nodes

--- base
+++ head
@@ -19,7 +19,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -44,7 +45,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[36] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[36] RotaryEmbedding: input_ids [65, 32, 33] → [64, 32, 33]
  • node[38] TensorScatter: input_ids [5, 67, 6] → [4, 70, 6]
qwen2_moe / model — 295 change(s)

Op summary: 214 → 224 nodes

--- base
+++ head
@@ -41,66 +41,71 @@
 Div
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -138,77 +143,82 @@
 Div
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Transpose
-MatMul
-Sigmoid
-Mul
-Add
-Add
-RMSNormalization
-Transpose
-MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Transpose
+MatMul
+Sigmoid
+Mul
+Add
+Add
+RMSNormalization
+Transpose
+MatMul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Add
  • + Add
  • + RMSNormalization
  • + Transpose
  • + MatMul

Removed nodes:

  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Transpose
  • - MatMul
  • - Sigmoid
  • - Mul
  • - Add
  • - Add
  • - RMSNormalization
  • - Transpose
  • - MatMul

Connectivity changes:

  • node[68] Mul: input_ids [135, 140] → [113, 140]
  • node[75] Mul: input_ids [145, 147] → [146, 147]
  • node[76] Transpose: input_ids [34] → [33]
  • node[77] MatMul: input_ids [148, 149] → [105, 149]
  • node[88] Transpose: input_ids [36] → [35]
  • node[108] Transpose: input_ids [41] → [39]
  • node[117] Add: input_ids [189, 44] → [176, 189]
  • node[149] Mul: input_ids [213, 224] → [222, 224]
  • node[157] Mul: input_ids [230, 232] → [227, 232]
  • node[158] Transpose: input_ids [57] → [55]
  • node[159] MatMul: input_ids [233, 234] → [210, 234]
  • node[172] Mul: input_ids [245, 247] → [242, 247]
  • node[180] Mul: input_ids [250, 255] → [253, 255]
  • node[193] Mul: input_ids [213, 268] → [267, 268]
  • node[197] Transpose: input_ids [64] → [63]
  • node[198] MatMul: input_ids [205, 273] → [272, 273]
  • node[202] Mul: input_ids [275, 277] → [218, 277]
qwen2_moe (static-cache) / model — 295 change(s)

Op summary: 204 → 214 nodes

--- base
+++ head
@@ -29,66 +29,71 @@
 Div
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -128,77 +133,82 @@
 Div
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Transpose
-MatMul
-Sigmoid
-Mul
-Add
-Add
-RMSNormalization
-Transpose
-MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Transpose
+MatMul
+Sigmoid
+Mul
+Add
+Add
+RMSNormalization
+Transpose
+MatMul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Add
  • + Add
  • + RMSNormalization
  • + Transpose
  • + MatMul

Removed nodes:

  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Transpose
  • - MatMul
  • - Sigmoid
  • - Mul
  • - Add
  • - Add
  • - RMSNormalization
  • - Transpose
  • - MatMul

Connectivity changes:

  • node[56] Mul: input_ids [119, 124] → [97, 124]
  • node[63] Mul: input_ids [129, 131] → [130, 131]
  • node[64] Transpose: input_ids [30] → [29]
  • node[65] MatMul: input_ids [132, 133] → [89, 133]
  • node[76] Transpose: input_ids [32] → [31]
  • node[96] Transpose: input_ids [37] → [35]
  • node[105] Add: input_ids [173, 40] → [160, 173]
  • node[139] Mul: input_ids [199, 210] → [208, 210]
  • node[147] Mul: input_ids [216, 218] → [213, 218]
  • node[148] Transpose: input_ids [53] → [51]
  • node[149] MatMul: input_ids [219, 220] → [196, 220]
  • node[162] Mul: input_ids [231, 233] → [228, 233]
  • node[170] Mul: input_ids [236, 241] → [239, 241]
  • node[183] Mul: input_ids [199, 254] → [253, 254]
  • node[187] Transpose: input_ids [60] → [59]
  • node[188] MatMul: input_ids [191, 259] → [258, 259]
  • node[192] Mul: input_ids [261, 263] → [204, 263]
qwen3 / model — 9 change(s)

Op summary: 72 → 74 nodes

--- base
+++ head
@@ -31,7 +31,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -60,7 +61,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[54] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[48] Reshape: input_ids [83, 15] → [82, 15]
  • node[50] RMSNormalization: input_ids [87, 29] → [87, 28]
qwen3 (static-cache) / model — 10 change(s)

Op summary: 68 → 70 nodes

--- base
+++ head
@@ -25,7 +25,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -56,7 +57,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish

Modified attributes:

  • node[48] RotaryEmbedding: num_heads: 2 → 4

Connectivity changes:

  • node[42] Reshape: input_ids [77, 15] → [76, 15]
  • node[44] RMSNormalization: input_ids [81, 29] → [81, 28]
  • node[50] TensorScatter: input_ids [5, 79, 6] → [4, 88, 6]
qwen3_5_moe (hybrid-text-generation) / model — 299 change(s)

Op summary: 264 → 275 nodes

--- base
+++ head
@@ -62,7 +62,8 @@
 Reshape
 RMSNormalization
 Cast
-Swish
+Sigmoid
+Mul
 Cast
 Mul
 CastLike
@@ -81,66 +82,71 @@
 Softmax
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -187,78 +193,83 @@
 Softmax
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Transpose
-MatMul
-Sigmoid
-Mul
-Add
-Add
-Add
-RMSNormalization
-Transpose
-MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Transpose
+MatMul
+Sigmoid
+Mul
+Add
+Add
+Add
+RMSNormalization
+Transpose
+MatMul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Add
  • + Add
  • + Add
  • + RMSNormalization
  • + Transpose
  • + MatMul

Removed nodes:

  • - Swish
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Transpose
  • - MatMul
  • - Sigmoid
  • - Mul
  • - Add
  • - Add
  • - Add
  • - RMSNormalization
  • - Transpose
  • - MatMul

Connectivity changes:

  • node[74] Add: input_ids [27, 16] → [75, 152]
  • node[94] Mul: input_ids [168, 173] → [161, 173]
  • node[100] Mul: input_ids [177, 179] → [178, 179]
  • node[101] Transpose: input_ids [34] → [33]
  • node[102] MatMul: input_ids [180, 181] → [155, 181]
  • node[113] Transpose: input_ids [36] → [35]
  • node[114] MatMul: input_ids [154, 193] → [155, 193]
  • node[151] Mul: input_ids [227, 230] → [228, 230]
  • node[171] Reshape: input_ids [250, 48] → [249, 51]
  • node[193] Transpose: input_ids [59] → [57]
  • node[194] MatMul: input_ids [276, 277] → [270, 277]
  • node[214] Mul: input_ids [292, 297] → [295, 297]
  • node[227] Mul: input_ids [270, 310] → [309, 310]
  • node[231] Transpose: input_ids [66] → [65]
  • node[232] MatMul: input_ids [264, 315] → [314, 315]
  • node[236] Mul: input_ids [317, 319] → [276, 319]
qwen3_5_text (hybrid-text-generation) / model — 18 change(s)

Op summary: 126 → 129 nodes

--- base
+++ head
@@ -62,7 +62,8 @@
 Reshape
 RMSNormalization
 Cast
-Swish
+Sigmoid
+Mul
 Cast
 Mul
 CastLike
@@ -76,7 +77,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -113,7 +115,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish
  • - Swish

Connectivity changes:

  • node[74] Add: input_ids [27, 16] → [47, 124]
  • node[91] Reshape: input_ids [142, 34] → [141, 33]
  • node[94] Transpose: input_ids [36] → [35]
  • node[95] MatMul: input_ids [137, 146] → [139, 146]
  • node[99] Add: input_ids [39, 16] → [38, 16]
  • node[100] RMSNormalization: input_ids [148, 151] → [143, 151]
  • node[118] Mul: input_ids [169, 171] → [170, 171]
  • node[119] Transpose: input_ids [44] → [43]
  • node[120] MatMul: input_ids [172, 173] → [168, 173]
qwen3_5_vl (hybrid-qwen-vl) / decoder — 19 change(s)

Op summary: 149 → 152 nodes

--- base
+++ head
@@ -84,7 +84,8 @@
 Reshape
 RMSNormalization
 Cast
-Swish
+Sigmoid
+Mul
 Cast
 Mul
 CastLike
@@ -98,7 +99,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -136,7 +138,8 @@
 RMSNormalization
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish
  • - Swish

Connectivity changes:

  • node[96] Add: input_ids [32, 21] → [0, 151]
  • node[108] Add: input_ids [36, 21] → [152, 163]
  • node[114] Reshape: input_ids [170, 39] → [169, 38]
  • node[117] Transpose: input_ids [41] → [40]
  • node[118] MatMul: input_ids [165, 174] → [167, 174]
  • node[122] Add: input_ids [44, 21] → [43, 21]
  • node[123] RMSNormalization: input_ids [176, 179] → [171, 179]
  • node[141] Mul: input_ids [197, 199] → [198, 199]
  • node[142] Transpose: input_ids [49] → [48]
  • node[143] MatMul: input_ids [200, 201] → [196, 201]
qwen3_moe / model — 236 change(s)

Op summary: 194 → 202 nodes

--- base
+++ head
@@ -44,51 +44,55 @@
 Div
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -131,64 +135,68 @@
 Div
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-CastLike
-Mul
-ReduceSum
-Mul
-Add
-Add
-RMSNormalization
-Transpose
-MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+CastLike
+Mul
+ReduceSum
+Mul
+Add
+Add
+RMSNormalization
+Transpose
+MatMul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + CastLike
  • + Mul
  • + ReduceSum
  • + Mul
  • + Add
  • + Add
  • + RMSNormalization
  • + Transpose
  • + MatMul

Removed nodes:

  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - CastLike
  • - Mul
  • - ReduceSum
  • - Mul
  • - Add
  • - Add
  • - RMSNormalization
  • - Transpose
  • - MatMul

Connectivity changes:

  • node[71] Mul: input_ids [130, 135] → [108, 135]
  • node[78] Mul: input_ids [140, 142] → [141, 142]
  • node[79] Transpose: input_ids [35] → [34]
  • node[80] MatMul: input_ids [143, 144] → [100, 144]
  • node[91] Transpose: input_ids [37] → [36]
  • node[109] Transpose: input_ids [42] → [40]
  • node[110] MatMul: input_ids [169, 174] → [173, 174]
  • node[115] Reshape: input_ids [178, 22] → [175, 19]
  • node[116] Reshape: input_ids [179, 22] → [177, 19]
  • node[124] Transpose: input_ids [47] → [45]
  • node[125] MatMul: input_ids [190, 191] → [188, 191]
  • node[156] Mul: input_ids [198, 223] → [221, 223]
  • node[166] Transpose: input_ids [56] → [54]
  • node[167] MatMul: input_ids [233, 234] → [194, 234]
  • node[180] Mul: input_ids [245, 247] → [242, 247]
  • node[188] Mul: input_ids [250, 255] → [253, 255]
qwen3_moe (static-cache) / model — 40 change(s)

Op summary: 184 → 192 nodes

--- base
+++ head
@@ -32,7 +32,8 @@
 Div
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -46,7 +47,8 @@
 Mul
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -61,7 +63,8 @@
 Add
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -76,7 +79,8 @@
 Add
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -121,7 +125,8 @@
 Div
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -135,7 +140,8 @@
 Mul
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -150,7 +156,8 @@
 Add
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -165,7 +172,8 @@
 Add
 Transpose
 MatMul
-Swish
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul

Removed nodes:

  • - Swish
  • - Swish
  • - Swish
  • - Swish
  • - Swish
  • - Swish
  • - Swish
  • - Swish

Connectivity changes:

  • node[59] Mul: input_ids [114, 119] → [92, 119]
  • node[66] Mul: input_ids [124, 126] → [125, 126]
  • node[67] Transpose: input_ids [31] → [30]
  • node[68] MatMul: input_ids [127, 128] → [84, 128]
  • node[79] Transpose: input_ids [33] → [32]
  • node[97] Transpose: input_ids [38] → [36]
  • node[98] MatMul: input_ids [153, 158] → [157, 158]
  • node[103] Reshape: input_ids [162, 18] → [159, 15]
  • node[104] Reshape: input_ids [163, 18] → [161, 15]
  • node[114] Transpose: input_ids [43] → [41]
  • node[115] MatMul: input_ids [176, 177] → [174, 177]
  • node[146] Mul: input_ids [184, 209] → [207, 209]
  • node[156] Transpose: input_ids [52] → [50]
  • node[157] MatMul: input_ids [219, 220] → [180, 220]
  • node[170] Mul: input_ids [231, 233] → [228, 233]
  • node[178] Mul: input_ids [236, 241] → [239, 241]
qwen3_next (hybrid-text-generation) / model — 615 change(s)

Op summary: 562 → 585 nodes

--- base
+++ head
@@ -62,7 +62,8 @@
 Reshape
 RMSNormalization
 Cast
-Swish
+Sigmoid
+Mul
 Cast
 Mul
 CastLike
@@ -83,70 +84,75 @@
 Div
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -205,7 +211,8 @@
 Reshape
 RMSNormalization
 Cast
-Swish
+Sigmoid
+Mul
 Cast
 Mul
 CastLike
@@ -226,70 +233,75 @@
 Div
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -348,7 +360,8 @@
 Reshape
 RMSNormalization
 Cast
-Swish
+Sigmoid
+Mul
 Cast
 Mul
 CastLike
@@ -369,70 +382,75 @@
 Div
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
 Transpose
 MatMul
 Mul
@@ -481,82 +499,87 @@
 Div
 Transpose
 MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Constant
-Equal
-Cast
-Mul
-ReduceSum
-Unsqueeze
-Mul
-Add
-Transpose
-MatMul
-Swish
-Transpose
-MatMul
-Mul
-Transpose
-MatMul
-Transpose
-MatMul
-Sigmoid
-Mul
-Add
-Add
-Add
-RMSNormalization
-Transpose
-MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Constant
+Equal
+Cast
+Mul
+ReduceSum
+Unsqueeze
+Mul
+Add
+Transpose
+MatMul
+Sigmoid
+Mul
+Transpose
+MatMul
+Mul
+Transpose
+MatMul
+Transpose
+MatMul
+Sigmoid
+Mul
+Add
+Add
+Add
+RMSNormalization
+Transpose
+MatMul

Added nodes:

  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Constant
  • + Equal
  • + Cast
  • + Mul
  • + ReduceSum
  • + Unsqueeze
  • + Mul
  • + Add
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Transpose
  • + MatMul
  • + Mul
  • + Transpose
  • + MatMul
  • + Transpose
  • + MatMul
  • + Sigmoid
  • + Mul
  • + Add
  • + Add
  • + Add
  • + RMSNormalization
  • + Transpose
  • + MatMul

Removed nodes:

  • - Swish
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Swish
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Swish
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Constant
  • - Equal
  • - Cast
  • - Mul
  • - ReduceSum
  • - Unsqueeze
  • - Mul
  • - Add
  • - Transpose
  • - MatMul
  • - Swish
  • - Transpose
  • - MatMul
  • - Mul
  • - Transpose
  • - MatMul
  • - Transpose
  • - MatMul
  • - Sigmoid
  • - Mul
  • - Add
  • - Add
  • - Add
  • - RMSNormalization
  • - Transpose
  • - MatMul

Modified attributes:

  • node[171] Transpose: perm: [0, 2, 1] → [1, 0]
  • node[178] Constant: value_float: None → 0.0, value_ints: [32, 32, 64] → None
  • node[320] Transpose: perm: [0, 2, 1] → [1, 0]
  • node[424] Constant: value_int: 3 → 2

Connectivity changes:

  • node[74] Add: input_ids [31, 20] → [135, 212]
  • node[103] Mul: input_ids [240, 242] → [241, 242]
  • node[104] Transpose: input_ids [38] → [37]
  • node[105] MatMul: input_ids [243, 244] → [215, 244]
  • node[112] Mul: input_ids [245, 251] → [223, 251]
  • node[117] Transpose: input_ids [40] → [39]
  • node[118] MatMul: input_ids [214, 257] → [215, 257]
  • node[157] Mul: input_ids [293, 296] → [294, 296]
  • node[169] Transpose: input_ids [55] → [52]
  • node[170] MatMul: input_ids [301, 309] → [307, 309]
  • node[171] Transpose: input_ids [304] → [53]
  • node[177] Transpose: input_ids [316] → [310]
  • node[226] Transpose: input_ids [61] → [60]
  • node[227] MatMul: input_ids [362, 371] → [369, 370]
  • node[252] Mul: input_ids [370, 396] → [395, 396]
  • node[255] Mul: input_ids [393, 399] → [397, 399]
  • node[273] Transpose: input_ids [70] → [69]
  • node[274] MatMul: input_ids [362, 418] → [417, 418]
  • node[278] Mul: input_ids [420, 422] → [377, 422]
  • node[320] Transpose: input_ids [464] → [81]
  • node[375] Transpose: input_ids [91] → [88]
  • node[376] MatMul: input_ids [524, 525] → [523, 524]
  • node[419] Transpose: input_ids [99] → [96]
  • node[420] MatMul: input_ids [510, 569] → [523, 569]
  • node[422] Transpose: input_ids [100] → [97]
  • node[425] Equal: input_ids [516, 574] → [529, 574]
  • node[427] Mul: input_ids [518, 576] → [531, 576]
  • node[431] Add: input_ids [565, 580] → [564, 580]
  • node[432] Transpose: input_ids [101] → [98]
  • node[433] MatMul: input_ids [510, 582] → [523, 582]
  • node[453] Transpose: input_ids [109] → [102]
  • node[454] MatMul: input_ids [597, 604] → [523, 603]
  • node[486] Mul: input_ids [637, 639] → [635, 638]
  • node[487] Transpose: input_ids [119] → [114]
  • node[488] MatMul: input_ids [640, 641] → [639, 640]
  • node[499] Transpose: input_ids [121] → [117]
  • node[500] MatMul: input_ids [626, 653] → [644, 653]
  • node[515] Transpose: input_ids [124] → [120]
  • node[516] MatMul: input_ids [626, 669] → [644, 669]
  • node[555] Mul: input_ids [705, 708] → [706, 708]
whisper (speech-to-text) / encoder — 2 change(s)

Op summary: 50 → 50 nodes

No op-sequence changes.

Modified attributes:

  • node[0] Conv: group: 1 → None
  • node[2] Conv: group: 1 → None

Legend: ⚪ No change · 🔵 Minor (attrs/inits) · 🟡 Moderate (nodes added/removed) · 🔴 Major (interface changed)

justinchuby and others added 6 commits August 15, 2026 21:38
@justinchuby
Make save progress callback thread-safe
0f0914d 
Serialize tqdm mutations and count callback invocations rather than relying
on index order. This supports concurrent callbacks from newer onnx_ir
versions while remaining compatible with the serial callback behavior in
onnx_ir 1.0.0.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
Show shard filename in save progress
0f1a1e6 
Include CallbackInfo.filename in the tqdm description so concurrent ONNX
shard saves visibly identify the shard currently reporting progress.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
Render one save progress bar per shard
ea27cf7 
Use CallbackInfo shard metadata from newer onnx_ir releases to create a
fixed-position tqdm bar for each external-data file. Fall back to one
global bar with onnx_ir 1.0.0, which does not expose per-shard counts.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
Use onnx_data suffix for external weights
10a30f7 
Name the default external data file model.onnx_data so sharding naturally
produces model-00001-of-00003.onnx_data rather than inserting the shard
suffix between .onnx and .data.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
Keep the conventional onnx.data filename
5da048e 
Restore model.onnx.data as the default external-data location. Newer
onnx_ir releases preserve this compound suffix when numbering shards;
the compatibility test also accepts the legacy 1.0.0 shard spelling.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: af83bc37-b9d6-4b94-be64-a4daa5a7ce40
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
@justinchuby
Order shard progress bars deterministically
e4ae9e9 
Derive each tqdm row from the one-based shard number embedded in the external-data filename, so concurrent callbacks cannot reorder the displayed bars. Cover shard-two-first callback delivery while retaining the onnx_ir 1.0 single-bar fallback.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

Copilot code review Copilot
Copilot review effort
Applies to this pull request for everyone.Learn more about Copilot code review.
Copilot left review comments

At least 1 approving review is required to merge this pull request.

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@justinchuby