[Qwen3.5] dedup position ids

[daijh]

Extract _make_synthetic_position_ids() to derive batch_size and sequence_length from the position_ids model input [3, B, S] instead of Shape on intermediate Q/K tensors. Use a fixed basename so make_node dedup creates the subgraph once and reuses it across all layers and Q/K calls.

• Reduces ONNX graph size by eliminating per-layer duplicates

• Eliminates redundant INT64 Mul nodes that cause CPU fallback on WebGPU (WebGPU Mul does not support INT64)

[Copilot]

Pull request overview

This PR reduces ONNX graph duplication in the Qwen3.5 builder by centralizing synthetic position_ids generation and deduplicating the resulting subgraph across all layers/attention paths, helping shrink graphs and avoid repeated INT64 ops.

Changes:

• Extracted _make_synthetic_position_ids() to derive batch_size/sequence_length from the position_ids model input shape [3, B, S].
• Switched mRoPE rotation to reuse a shared synthetic position_ids subgraph via a fixed basename (leveraging name-based node dedup).
• Removed per-Q/K (and per-layer) synthetic position_ids subgraph construction from _apply_mrope_rotation().

[daijh]

• Total Nodes: 1509 → 1434 (75 nodes pruned)
• WebGPU-EP MemcpyFromHost: 16 → 5 (11 host-to-device transfers eliminated)

Performance Comparison - Prefill-1024, Decode-128

Model Variant	Prefill Speed (TPS)	Decode Speed (TPS)
Qwen3.5-0.8B-int4	2415.4	56.5
Qwen3.5-0.8B-int4-with-optimization	2452.8	58.4

[daijh]

@apsonawane @kunal-vaishnavi PTAL

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

There are some styling changes that can be made.

[apsonawane]

Lgtm after Kunal's and Copilot changes are done

[daijh]

Thanks.
Updated the PRs to address all the comments.

[daijh]

When I run lintrunner -a, it's touching unrelated code.
Should I include these lint fixes in this PR, or leave them out for now?

diff --git a/src/python/py/models/builders/qwen.py b/src/python/py/models/builders/qwen.py
index 4e47e5c5..ffabd1ee 100644
--- a/src/python/py/models/builders/qwen.py
+++ b/src/python/py/models/builders/qwen.py
@@ -449,9 +449,7 @@ class Qwen25VLTextModel(Model):
 
         seq_len_node = f"{basename}/SeqLen/Gather"
         seq_len_out = f"{seq_len_node}/output_0"
-        self.make_gather(
-            seq_len_node, [f"{shape_node}/output_0", "/model/constants/INT64/1"], ir.DataType.INT64, [], 0
-        )
+        self.make_gather(seq_len_node, [f"{shape_node}/output_0", "/model/constants/INT64/1"], ir.DataType.INT64, [], 0)
 
         # Calculate Total Tokens = B * S
         mul_len_node = f"{basename}/TotalLen/Mul"
@@ -463,7 +461,10 @@ class Qwen25VLTextModel(Model):
         range_node = f"{basename}/Range"
         range_out = f"{range_node}/output_0"
         self.make_range(
-            range_node, ["/model/constants/INT64/0", mul_len_out, "/model/constants/INT64/1"], ir.DataType.INT64, ["total_token_count"]
+            range_node,
+            ["/model/constants/INT64/0", mul_len_out, "/model/constants/INT64/1"],
+            ir.DataType.INT64,
+            ["total_token_count"],
         )
         range_out = f"{range_node}/output_0"
 
@@ -1801,7 +1802,10 @@ class Qwen35TextModel(Model):
 
         a_plus_dt_name = f"{basename}/decay/Add"
         self.make_add(
-            a_plus_dt_name, [f"{a_cast_name}/output_0", dt_bias_init], ir.DataType.FLOAT, ["batch_size", "sequence_length", n_kv]
+            a_plus_dt_name,
+            [f"{a_cast_name}/output_0", dt_bias_init],
+            ir.DataType.FLOAT,
+            ["batch_size", "sequence_length", n_kv],
         )
         a_plus_dt_output = f"{a_plus_dt_name}/output_0"
 
@@ -1810,7 +1814,9 @@ class Qwen35TextModel(Model):
         softplus_output = f"{softplus_name}/output_0"
 
         g_fp32_name = f"{basename}/decay/Mul"
-        self.make_mul(g_fp32_name, [neg_exp_a_name, softplus_output], ir.DataType.FLOAT, ["batch_size", "sequence_length", n_kv])
+        self.make_mul(
+            g_fp32_name, [neg_exp_a_name, softplus_output], ir.DataType.FLOAT, ["batch_size", "sequence_length", n_kv]
+        )
         g_fp32_output = f"{g_fp32_name}/output_0"
 
         # Cast decay back to io_dtype for the kernel
@@ -1900,8 +1906,13 @@ class Qwen35TextModel(Model):
         self.make_mul(sq_name, [input_name, input_name], self.io_dtype, full_shape)
 
         sum_name = f"{basename}/SumSq/ReduceSum"
-        self.make_reduce_sum(sum_name, [f"{sq_name}/output_0", "/model/constants/INT64/[-1]"],
-                             self.io_dtype, reduced_shape, keepdims=True)
+        self.make_reduce_sum(
+            sum_name,
+            [f"{sq_name}/output_0", "/model/constants/INT64/[-1]"],
+            self.io_dtype,
+            reduced_shape,
+            keepdims=True,
+        )
 
         # sum(x^2) + eps
         eps_name = self._get_shared_l2_eps()
@@ -1990,12 +2001,17 @@ class Qwen35TextModel(Model):
         # output = norm * silu(z) in fp32
         gated_fp32_name = f"{basename}/gated_fp32/Mul"
         self.make_mul(
-            gated_fp32_name, [f"{norm_cast_name}/output_0", z_silu_output], ir.DataType.FLOAT, ["batch_size", "sequence_length", v_dim]
+            gated_fp32_name,
+            [f"{norm_cast_name}/output_0", z_silu_output],
+            ir.DataType.FLOAT,
+            ["batch_size", "sequence_length", v_dim],
         )
 
         # Cast back to io_dtype
         gated_name = f"{basename}/gated/Cast"
-        self.make_cast(gated_name, f"{gated_fp32_name}/output_0", self.io_dtype, ["batch_size", "sequence_length", v_dim])
+        self.make_cast(
+            gated_name, f"{gated_fp32_name}/output_0", self.io_dtype, ["batch_size", "sequence_length", v_dim]
+        )
         gated_output = f"{gated_name}/output_0"
 
         return gated_output

[daijh]

When I run lintrunner -a, it's touching unrelated code.
Should I include these lint fixes in this PR, or leave them out for now?

OKAY. Rebasing against main cleared the lint errors.

[daijh]

@kunal-vaishnavi
It looks like the previous CI failures were due to infrastructure issues.
Could you please re-trigger the checks and merge the PR if they pass?

[kunal-vaishnavi]

@kunal-vaishnavi It looks like the previous CI failures were due to infrastructure issues. Could you please re-trigger the checks and merge the PR if they pass?

You will need to sync with the main branch to get this CI fix.

[daijh]

@kunal-vaishnavi It looks like the previous CI failures were due to infrastructure issues. Could you please re-trigger the checks and merge the PR if they pass?

You will need to sync with the main branch to get this CI fix.

Rebase on main to include #2122.

[xiaofeihan1]

It seems like this PR is ready to merge, could we merge it? @daijh @kunal-vaishnavi

[daijh]

It seems like this PR is ready to merge, could we merge it? @daijh @kunal-vaishnavi

Thanks!
I don't have merge permissions as a non-committer.
@kunal-vaishnavi, could you please help land this?