[Intel GPU] Enable DeepSeek V4 Inference on XPU#25336

Merged

mingfeima merged 25 commits into

sgl-project:intel_devfrom

polisettyvarma:dev/jvarma/dsv4_xpu_intel_dev

May 18, 2026

Commits on May 14, 2026

feat: SM120 (Blackwell Desktop) support for DeepSeek-V4 inference

authored and
polisettyvarma
committed
Add torch fallback for CompressorPrefillPlan.generate when tvm_ffi is missing

rahulvijayaraghavan
authored and
polisettyvarma
committed
DeepSeek V4 - add triton and torch fallback for hc_split_sinkhorn
polisettyvarma
committed
Add triton fallback for fused_rope

rahulvijayaraghavan
authored and
polisettyvarma
committed
Support MXFP4 in MoE using fused_moe triton
polisettyvarma
committed
nsa_indexer: pure-torch Hadamard fallback for XPU

rahulvijayaraghavan
authored and
polisettyvarma
committed
Add torch/triton fallback for compressor kernels

rahulvijayaraghavan
authored and
polisettyvarma
committed
Add torch fallback for fused norm rope jit kernel

rahulvijayaraghavan
authored and
polisettyvarma
committed
support SGLANG_FP8_PAGED_MQA_LOGITS_TRITON also
polisettyvarma
committed
xpu: chunk SM120 sparse decode fallback to bound peak memory and clamp OOB indices

rahulvijayaraghavan
authored and
polisettyvarma
committed
topk: drop dead allocations in fused_topk_torch_native

rahulvijayaraghavan
authored and
polisettyvarma
committed
indexer: remove L0 sync hot spots in topk_transform_512_pytorch_vectorized

rahulvijayaraghavan
authored and
polisettyvarma
committed
flash_mla_sm120_fallback: hoist SGLANG_SM120_SPARSE_CHUNK_MIB to import time

rahulvijayaraghavan
authored and
polisettyvarma
committed
jit_kernel/deepseek_v4: remove L0 host-sync hot spots in torch fallbacks

rahulvijayaraghavan
authored and
polisettyvarma
committed
optimize _torch_plan_compress_prefill memory usage
polisettyvarma
committed
Add sqrtsoftplus support to fused_topk_torch_native

rahulvijayaraghavan
authored and
polisettyvarma
committed
Make tilelang import optional in deepseek_v4_rope

rahulvijayaraghavan
authored and
polisettyvarma
committed
changes required to pass flashmla metadata and flashmla xpu spport
polisettyvarma
committed
add pytorch profiling support for XPU in bench offline throughput
polisettyvarma
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Intel GPU] Enable DeepSeek V4 Inference on XPU#25336

[Intel GPU] Enable DeepSeek V4 Inference on XPU#25336
mingfeima merged 25 commits into
sgl-project:intel_devfrom
polisettyvarma:dev/jvarma/dsv4_xpu_intel_dev

Commits on May 14, 2026

feat: SM120 (Blackwell Desktop) support for DeepSeek-V4 inference

Add torch fallback for CompressorPrefillPlan.generate when tvm_ffi is missing

DeepSeek V4 - add triton and torch fallback for hc_split_sinkhorn

Add triton fallback for fused_rope

Support MXFP4 in MoE using fused_moe triton

nsa_indexer: pure-torch Hadamard fallback for XPU

Add torch/triton fallback for compressor kernels

Add torch fallback for fused norm rope jit kernel

support SGLANG_FP8_PAGED_MQA_LOGITS_TRITON also

xpu: chunk SM120 sparse decode fallback to bound peak memory and clamp OOB indices

topk: drop dead allocations in fused_topk_torch_native

indexer: remove L0 sync hot spots in topk_transform_512_pytorch_vectorized

flash_mla_sm120_fallback: hoist SGLANG_SM120_SPARSE_CHUNK_MIB to import time

jit_kernel/deepseek_v4: remove L0 host-sync hot spots in torch fallbacks

optimize _torch_plan_compress_prefill memory usage

Add sqrtsoftplus support to fused_topk_torch_native

Make tilelang import optional in deepseek_v4_rope

changes required to pass flashmla metadata and flashmla xpu spport

add pytorch profiling support for XPU in bench offline throughput

Commits on May 15, 2026

fix bug in torch compress plan prefill due to conflict resolving incorrectly

remove _torch_plan_compress_prefill implementation

remove _triton_compress_forward implementation

remove unnecessary changes

remove _hc_split_sinkhorn_torch implementation

iset default env for XPU