pytorch - ✅(Solved) Fix Fusable GQA-style QKV matmuls are overlooked [1 pull requests, 2 comments, 2 participants]

pytorch2026-03-25 14:17:04

ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

GitHub issue URL

Helpful · Quick feedback

GitHub stats

pytorch/pytorch#178387•Fetched 2026-04-08 01:30:44

View on GitHub

Comments

Participants

Timeline

Reactions

Author

Participants

Assignees

Timeline (top)

subscribed ×31mentioned ×30labeled ×8referenced ×6

Fix Action

Fix / Workaround

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 48 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 32 On-line CPU(s) list: 0-31 Vendor ID: AuthenticAMD Model name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S CPU family: 26 Model: 112 Thread(s) per core: 2 Core(s) per socket: 16 Socket(s): 1 Stepping: 0 Frequency boost: enabled CPU(s) scaling MHz: 50% CPU max MHz: 5187.0000 CPU min MHz: 599.0000 BogoMIPS: 5988.24 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good amd_lbr_v2 nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk avx_vnni avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid bus_lock_detect movdiri movdir64b overflow_recov succor smca fsrm avx512_vp2intersect flush_l1d amd_lbr_pmc_freeze Virtualization: AMD-V L1d cache: 768 KiB (16 instances) L1i cache: 512 KiB (16 instances) L2 cache: 16 MiB (16 instances) L3 cache: 64 MiB (2 instances) NUMA node(s): 1 NUMA node0 CPU(s): 0-31 Vulnerability Gather data sampling: Not affected Vulnerability Ghostwrite: Not affected Vulnerability Indirect target selection: Not affected Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec rstack overflow: Mitigation; IBPB on VMEXIT only Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsa: Not affected Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Mitigation; IBPB on VMEXIT

PR fix notes

PR #178523: Relax concat-linear fusion to support GQA QKV

Repository: pytorch/pytorch
Author: eble-amd
State: closed | merged: False
Link: https://github.com/pytorch/pytorch/pull/178523

Description (problem / solution / changelog)

The freezing concat-linear optimization fuses independent matmuls that share the same input (e.g., QKV projections) into a single GEMM. It previously required all weights to have identical shapes, preventing fusion of GQA-style attention where Q has a larger output dimension than K/V (e.g., Q=960, K=V=320).

Relax the shape check to require only that non-concatenated dimensions match (shape[:-1]), allowing weights and biases to differ in the last dimension. Use split (which supports unequal sizes) instead of chunk (which requires equal sizes) to split the fused result.

Fixes #178387, "Fusable GQA-style QKV matmuls are overlooked."

Using the reproducer script from the ticket with commit c0c089817e69e999ee07b5efa0396c1cc5b9c655,

$ LD_LIBRARY_PATH="$(rocm-sdk path --root)/lib:${LD_LIBRARY_PATH}" python ./repro_concat_linear.py
PASS: 1 matmul(s) in generated code

cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo

Changed files

test/inductor/test_inductor_freezing.py (modified, +2/-4)
torch/_inductor/fx_passes/freezing_patterns.py (modified, +10/-7)

Code Example

"""Reproducer: does TorchInductor concat-linear fuse GQA-style QKV projections?
The submitter ran this in a ROCm venv:

    LD_LIBRARY_PATH="$(rocm-sdk path --root)/lib:${LD_LIBRARY_PATH}" \
        python ./repro_concat_linear.py
"""
import re
import torch
import torch._inductor.config as inductor_config
from torch._inductor.utils import run_and_get_code

inductor_config.freezing = True

class GQA_QKV(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.q_proj = torch.nn.Linear(720, 960, bias=False)
        self.k_proj = torch.nn.Linear(720, 320, bias=False)
        self.v_proj = torch.nn.Linear(720, 320, bias=False)

    def forward(self, x):
        return self.q_proj(x), self.k_proj(x), self.v_proj(x)

model = GQA_QKV().cuda().bfloat16().eval()
x = torch.randn(50, 720, device="cuda", dtype=torch.bfloat16)

compiled = torch.compile(model, mode="max-autotune-no-cudagraphs")
with torch.no_grad():
    _, code = run_and_get_code(compiled, x)

# Extract the "def call" body from the generated code and count mm kernel
# launches.  Fully fused: 1 launch.  Partially fused (pre-existing): 2.
call_body = re.search(r"def call\(self.*?return .*?\n", code[0], re.DOTALL).group()
mm_calls = re.findall(r"\.run\(", call_body)
fused = len(mm_calls) == 1
print(f"{'PASS' if fused else 'FAIL'}: {len(mm_calls)} matmul(s) in generated code")

---

FAIL: 3 matmul(s) in generated code

---

PASS: 1 matmul(s) in generated code

---

Collecting environment information...
PyTorch version: 2.12.0a0+gitb19e5a7
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 7.13.60800

OS: Ubuntu 24.04.3 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version: Could not collect
CMake version: version 3.31.10
Libc version: glibc-2.39

Python version: 3.12.3 (main, Jan  8 2026, 11:30:50) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.14.0-1019-oem-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: 
GPU models and configuration: Radeon 8060S Graphics (gfx1151)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
Is XPU available: False
HIP runtime version: 7.13.60800
MIOpen runtime version: 3.5.1
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           48 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  32
On-line CPU(s) list:                     0-31
Vendor ID:                               AuthenticAMD
Model name:                              AMD RYZEN AI MAX+ 395 w/ Radeon 8060S
CPU family:                              26
Model:                                   112
Thread(s) per core:                      2
Core(s) per socket:                      16
Socket(s):                               1
Stepping:                                0
Frequency boost:                         enabled
CPU(s) scaling MHz:                      50%
CPU max MHz:                             5187.0000
CPU min MHz:                             599.0000
BogoMIPS:                                5988.24
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good amd_lbr_v2 nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk avx_vnni avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid bus_lock_detect movdiri movdir64b overflow_recov succor smca fsrm avx512_vp2intersect flush_l1d amd_lbr_pmc_freeze
Virtualization:                          AMD-V
L1d cache:                               768 KiB (16 instances)
L1i cache:                               512 KiB (16 instances)
L2 cache:                                16 MiB (16 instances)
L3 cache:                                64 MiB (2 instances)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-31
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; IBPB on VMEXIT only
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB on VMEXIT

Versions of relevant libraries:
[pip3] numpy==2.4.3
[pip3] optree==0.19.0
[pip3] torch==2.12.0a0+gitb19e5a7
[pip3] triton==3.6.0+rocm7.13.0a20260324
[conda] Could not collect

RAW_BUFFERClick to expand / collapse

🐛 Describe the bug

TorchInductor's freezing_patterns.py has a concat-linear optimization that fuses independent matmuls sharing the same input (e.g. QKV projections, gate+up in MLP) into a single larger GEMM. This reduces kernel launch overhead, which is critical for small-GEMM-dominated VLA models like SmolVLA on Strix Halo (gfx1151).

Currently fused (when inductor_config.freezing = True):

gate+up projections with same output dim (e.g. 2 x (720->2048) > 1 x (720>4096))
QKV where all three have the same output dim (e.g. 3 x (768->768) > 1 x (768>2304))

Not fused due to same-shape requirement in check_concat_weights():

GQA-style QKV where Q has different output dim than K/V (e.g. Q=960, K=320, V=320)

"""Reproducer: does TorchInductor concat-linear fuse GQA-style QKV projections?
The submitter ran this in a ROCm venv:

    LD_LIBRARY_PATH="$(rocm-sdk path --root)/lib:${LD_LIBRARY_PATH}" \
        python ./repro_concat_linear.py
"""
import re
import torch
import torch._inductor.config as inductor_config
from torch._inductor.utils import run_and_get_code

inductor_config.freezing = True

class GQA_QKV(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.q_proj = torch.nn.Linear(720, 960, bias=False)
        self.k_proj = torch.nn.Linear(720, 320, bias=False)
        self.v_proj = torch.nn.Linear(720, 320, bias=False)

    def forward(self, x):
        return self.q_proj(x), self.k_proj(x), self.v_proj(x)

model = GQA_QKV().cuda().bfloat16().eval()
x = torch.randn(50, 720, device="cuda", dtype=torch.bfloat16)

compiled = torch.compile(model, mode="max-autotune-no-cudagraphs")
with torch.no_grad():
    _, code = run_and_get_code(compiled, x)

# Extract the "def call" body from the generated code and count mm kernel
# launches.  Fully fused: 1 launch.  Partially fused (pre-existing): 2.
call_body = re.search(r"def call\(self.*?return .*?\n", code[0], re.DOTALL).group()
mm_calls = re.findall(r"\.run\(", call_body)
fused = len(mm_calls) == 1
print(f"{'PASS' if fused else 'FAIL'}: {len(mm_calls)} matmul(s) in generated code")

Output at b19e5a7f7329dadcf639fd6fc53f934588193595:

FAIL: 3 matmul(s) in generated code

Expected output:

PASS: 1 matmul(s) in generated code

Versions

Environment gathered with curl -sL https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py | LD_LIBRARY_PATH="$(rocm-sdk path --root)/lib:${LD_LIBRARY_PATH}" python.

Collecting environment information...
PyTorch version: 2.12.0a0+gitb19e5a7
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 7.13.60800

OS: Ubuntu 24.04.3 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version: Could not collect
CMake version: version 3.31.10
Libc version: glibc-2.39

Python version: 3.12.3 (main, Jan  8 2026, 11:30:50) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.14.0-1019-oem-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: 
GPU models and configuration: Radeon 8060S Graphics (gfx1151)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
Is XPU available: False
HIP runtime version: 7.13.60800
MIOpen runtime version: 3.5.1
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           48 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  32
On-line CPU(s) list:                     0-31
Vendor ID:                               AuthenticAMD
Model name:                              AMD RYZEN AI MAX+ 395 w/ Radeon 8060S
CPU family:                              26
Model:                                   112
Thread(s) per core:                      2
Core(s) per socket:                      16
Socket(s):                               1
Stepping:                                0
Frequency boost:                         enabled
CPU(s) scaling MHz:                      50%
CPU max MHz:                             5187.0000
CPU min MHz:                             599.0000
BogoMIPS:                                5988.24
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good amd_lbr_v2 nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local user_shstk avx_vnni avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid bus_lock_detect movdiri movdir64b overflow_recov succor smca fsrm avx512_vp2intersect flush_l1d amd_lbr_pmc_freeze
Virtualization:                          AMD-V
L1d cache:                               768 KiB (16 instances)
L1i cache:                               512 KiB (16 instances)
L2 cache:                                16 MiB (16 instances)
L3 cache:                                64 MiB (2 instances)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-31
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; IBPB on VMEXIT only
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB on VMEXIT

Versions of relevant libraries:
[pip3] numpy==2.4.3
[pip3] optree==0.19.0
[pip3] torch==2.12.0a0+gitb19e5a7
[pip3] triton==3.6.0+rocm7.13.0a20260324
[conda] Could not collect

cc @jerryzh168 @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @aakhundov @coconutruben

extent analysis

Fix Plan

To fix the issue of TorchInductor's freezing_patterns.py not fusing GQA-style QKV projections, we need to modify the check_concat_weights() function to relax the same-shape requirement.

Here are the steps:

Modify the check_concat_weights() function to allow concatenation of weights with different shapes.
Update the concat_linear() function to handle the new weights concatenation logic.

Code Changes

def check_concat_weights(weights):
    # Relax the same-shape requirement
    return all(weight.dtype == weights[0].dtype for weight in weights)

def concat_linear(linears):
    weights = [linear.weight for linear in linears]
    if check_concat_weights(weights):
        # Concatenate the weights
        concatenated_weight = torch.cat(weights, dim=0)
        # Create a new linear layer with the concatenated weight
        new_linear = torch.nn.Linear(linears[0].in_features, concatenated_weight.shape[0], bias=False)
        new_linear.weight = concatenated_weight
        return new_linear
    else:
        # Handle the case where weights cannot be concatenated
        return linears

Verification

To verify that the fix worked, we can use the provided reproducer code:

class GQA_QKV(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.q_proj = torch.nn.Linear(720, 960, bias=False)
        self.k_proj = torch.nn.Linear(720, 320, bias=False)
        self.v_proj = torch.nn.Linear(720, 320, bias=False)

    def forward(self, x):
        return self.q_proj(x), self.k_proj(x), self.v_proj(x)

model = GQA_QKV().cuda().bfloat16().eval()
x = torch.randn(50, 720, device="cuda", dtype=torch.bfloat16)

compiled = torch.compile(model, mode="max-autotune-no-cudagraphs")
with torch.no_grad():
    _, code = run_and_get_code(compiled, x)

call_body = re.search(r"def call\(self.*?return .*?\n", code[0], re.DOTALL).group()
mm_calls = re.findall(r"\.run\(", call_body)
fused = len(mm_calls) == 1
print(f"{'PASS' if fused else 'FAIL'}: {len(mm_calls)} matmul(s) in generated code")

If the fix is correct, the output should be PASS: 1 matmul(s) in generated code.

Extra Tips

Make sure to test the fix with different input shapes and sizes to ensure that it works correctly in all cases.
Consider adding additional tests to the TorchInductor test suite to cover this specific scenario.
If you encounter any issues or regressions, please report them to the TorchInductor maintainers.

Vote matrix · Quick signals

Works

Did the solution work? Tap to confirm.

Easy Fix

Was it a quick fix?

Time Saver

Did it save you time?

Blocking

Was it severely blocking?

Common Issue

Are others likely hitting this too?

Flaky / Intermittent

Is it intermittent?

Verified / Reproducible

Can you reproduce it reliably?

#api #optimization #retrieval issue #search optimization #API routing

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Data

Security

Network

Code

UI/UX

Text

System

Multimedia

Protocol

API

Engineering

pytorch - ✅(Solved) Fix Fusable GQA-style QKV matmuls are overlooked [1 pull requests, 2 comments, 2 participants]

Recommended Tools

GitHub issue graph ai analysis

Fix Action

Fix / Workaround

PR fix notes

PR #178523: Relax concat-linear fusion to support GQA QKV

Description (problem / solution / changelog)

Changed files

Code Example

🐛 Describe the bug

Versions

extent analysis

Fix Plan

Code Changes

Verification

Extra Tips

Still need to ship something?

TRENDING

pytorch - ✅(Solved) Fix Fusable GQA-style QKV matmuls are overlooked [1 pull requests, 2 comments, 2 participants]

Recommended Tools

GitHub issue graph ai analysis

Fix Action

Fix / Workaround

PR fix notes

PR #178523: Relax concat-linear fusion to support GQA QKV

Description (problem / solution / changelog)

Changed files

Code Example

🐛 Describe the bug

Versions

extent analysis

Fix Plan

Code Changes

Verification

Extra Tips

Still need to ship something?

RELATED_DISCOVERY

TRENDING