vllm.models.deepseek_v41.nvidia.specialized_decode.ampere.quantize
¶
The FP4 quantizers of DeepSeek-V4.1 on SM80, in CuTe DSL.
Three per-token kernels write the FP4 data that the stock SM100 kernels
produce with conversion instructions SM80 lacks (codecs.py encodes with
comparisons and integer operations instead). Each follows its
reference function, which states the rounding:
IndexerQuery-- the indexer queries: RoPE on dims 64-127, then MXFP4 (one UE8M0 scale per 32 dims), and the head weights times the softmax and head scales (reference.indexer_q_rope_quant). A warp per (token, head), four dims per lane.IndexKey-- the index keys of group-boundary latents: RMSNorm, RoPE at the group's first position, MXFP4, stored into the paged key cache (reference.indexer_k_norm_ropeandstore_index_keys). A warp per token, four dims per lane.CompressedRecord-- the 288-byte NVFP4 compressed KV record of a group-boundary latent (reference.nvfp4_kv_record), stored into the compressed cache. A warp per token, one 16-dim scale tile per lane.
The two storing kernels skip rows whose slot is negative or whose position does not close a group. Byte outputs are written by address.
Classes:
-
CompressedRecord–NVFP4 compressed KV records into the paged compressed cache; see the
-
IndexKey–MXFP4 index keys into the paged key cache; see the module docstring.
-
IndexerQuery–MXFP4 indexer queries and scaled head weights; see the module
Functions:
-
indexer_q_rope_quant–reference.indexer_q_rope_quanton the GPU. -
store_compressed_records–reference.nvfp4_kv_recordof the group-boundary latents, stored at -
store_index_keys–reference.indexer_k_norm_ropethenstore_index_keys, fused.
CompressedRecord
¶
NVFP4 compressed KV records into the paged compressed cache; see the module docstring.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 | |
IndexKey
¶
MXFP4 index keys into the paged key cache; see the module docstring.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 | |
IndexerQuery
¶
MXFP4 indexer queries and scaled head weights; see the module docstring.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 | |
_cache_geometry(cache, record)
¶
(block size, block stride in bytes) of a uint8 paged cache whose
pages hold block size rows of record bytes.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
_fake(dtype, align)
¶
A placeholder pointer for compilation.
_key_kernel(device_index, slot_dtype)
cached
¶
IndexKey compiled for a device and slot-mapping dtype.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
_mxfp4_lane(x)
¶
MXFP4 of this lane's four values of a 32-value block spread over eight adjacent lanes: the two packed bytes as one 16-bit word (even element in the low nibble) and the block's UE8M0 scale byte.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
_pointer(dtype, tensor, align)
¶
A CuTe DSL global-memory pointer to tensor's data.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
_query_kernel(device_index, weights_dtype)
cached
¶
IndexerQuery compiled for a device and head-weight dtype.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
_record_kernel(device_index, slot_dtype)
cached
¶
CompressedRecord compiled for a device and slot-mapping dtype.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
_rope(x, i, c, s)
¶
GPT-J RoPE of the pair x[2i], x[2i + 1] in place.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
indexer_q_rope_quant(positions, q, cos_sin_cache, weights, weights_scale)
¶
reference.indexer_q_rope_quant on the GPU.
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
store_compressed_records(latent, positions, slot_mapping, cos_sin_cache, compress_ratio, kv_cache)
¶
reference.nvfp4_kv_record of the group-boundary latents, stored at
their slots of the compressed cache [blocks, block_size, 288].
Source code in vllm/models/deepseek_v41/nvidia/specialized_decode/ampere/quantize.py
store_index_keys(k_pre, positions, slot_mapping, norm_weight, eps, cos_sin_cache, compress_ratio, k_cache)
¶
reference.indexer_k_norm_rope then store_index_keys, fused.