[DeepEP] support MXFP4 quant

[zuje123]

DeepEP supports MXFP4 quantization (for 910D)

Test Results
================ test_low_latency.py ================
[rank 0] raw_num_tokens: 16, aligned_num_tokens: 16
[rank 1] raw_num_tokens: 16, aligned_num_tokens: 16
[testing] Running with quant_type='mxfp4', data=rand ...
[test] rank=0, low_latency_dispatch, simulated_gemm_x.shape=torch.Size([256, 7168]), simulated_gemm_x.dtype=torch.bfloat16
[test] rank=1, low_latency_dispatch, simulated_gemm_x.shape=torch.Size([256, 7168]), simulated_gemm_x.dtype=torch.bfloat16
[test] rank=0, low_latency_dispatch, packed_recv_count.shape=torch.Size([8]), packed_recv_count=tensor([15, 15, 20, 16, 15, 15, 18, 9], device='npu:0')
[test] rank=0, low_latency_dispatch, packed_recv_x[0].dtype=torch.float4_e2m1fn_x2, packed_recv_x[0].shape=torch.Size([256, 3584])
[test] rank=0, low_latency_dispatch, packed_recv_x[1].dtype=torch.float8_e8m0fnu, packed_recv_x[1].shape=torch.Size([57344])
[test] rank=1, low_latency_dispatch, packed_recv_count.shape=torch.Size([8]), packed_recv_count=tensor([22, 16, 18, 15, 16, 11, 16, 19], device='npu:1')
[test] rank=1, low_latency_dispatch, packed_recv_x[0].dtype=torch.float4_e2m1fn_x2, packed_recv_x[0].shape=torch.Size([256, 3584])
[test] rank=1, low_latency_dispatch, packed_recv_x[1].dtype=torch.float8_e8m0fnu, packed_recv_x[1].shape=torch.Size([57344])
[rank1]:[W623 05:25:23.427888390 compiler_depend.ts:41] Warning: Device do not support double dtype now, dtype cast replace with float. (function operator())
[rank0]:[W623 05:25:23.428825150 compiler_depend.ts:41] Warning: Device do not support double dtype now, dtype cast replace with float. (function operator())
rank 0 PASSED [quant_type='mxfp4'] avg_diff=0.00004, max_diff=1.98265, cosine_diff=0.00775
rank 1 PASSED [quant_type='mxfp4'] avg_diff=-0.00031, max_diff=1.98790, cosine_diff=0.00768
passed

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!